Abstract
We develop a theory for the temporal integration of visual motion motivated by psychophysical experiments. The theory proposes that input data are temporally grouped and used to predict and estimate the motion flows in the image sequence. This temporal grouping can be considered a generalization of the data association techniques used by engineers to study motion sequences. Our temporalgrouping theory is expressed in terms of the Bayesian generalization of standard Kalman filtering. To implement the theory we derive a parallel network which shares some properties of cortical networks. Computer simulations of this network demonstrate that our theory qualitatively accounts for psychophysical experiments on motion occlusion and motion outliers.
1 Introduction
Local motion signals are often ambiguous and many important motion phenomena can be explained by hypothesising that the human visual system uses temporal coherence to resolve ambiguous inputs^{1}^{1}1Temporal coherence is the assumption that motion in natural images is mostly temporally smooth, that is, it rarely changes direction and/or speed abruptly.. For example, the perceptual tendency to disambiguate the trajectory of an ambiguous motion path by using the time average of its past motion was first reported by Anstis and Ramachandran (Ramachandran and Anstis 1983; Anstis and Ramachandran 1987). They called this phenomenon motion inertia. Other motion phenomena involving temporal coherence include the improvement of velocity estimation over time (McKee et al. 1986, Ascher 1998), blur removal (Burr et al.
1986), motionoutlier detection (Watamaniuk
et al. 1994), and motion occlusion (Watamaniuk and McKee 1995). These motion phenomena pose a serious challenge to motionperception models. For example, in motionoutlier detection the target dot is indistinguishable from the background noise if one only observes a few frames of the motion. Therefore, the only way to detect the target dot is by means of its extended motion over time. Consequently, explanations based solely on single, large localmotion detectors as in the MotionEnergy and ElaboratedReichardt models (Hassenstein and Reichardt 1956; van Santen and Sperling 1984; Adelson and Bergen 1985; Watson and Ahumada 1985; for a review of motion models, see Nakayama 1985) seem inadequate to explain these phenomena (Verghese et al. 1999). Also, it has been argued (Yuille and Grzywacz 1998) that all these experiments could be interpreted in terms of temporal grouping of motion signals involving prediction, observation and estimation. The detection of targets, and their temporal grouping, could be achieved by verifying that the observations were consistent with the motion predictions. Conversely, failure in these predictions would indicate that the observations were due to noise, or distractors, which could be ignored. The ability of a system to predict the target’s motions thus represents a powerful mechanism to extract targets from background distractors and might be a powerful cue for the interpretation of visual scenes.There has been little theoretical work on temporal coherence. Grzywacz et al. (1989) developed a theoretical model which could account for motioninertia phenomena by requiring that the direction of motion varies slowly over time while speed could vary considerably faster. More recently Grzywacz et al. (1995) proposed a biologically plausible model that extends the work on motion inertia and allow for the detection of motion outliers. It was observed (Grzywacz et al 1989) that for general threedimensional motion, the direction, but not the speed, of the image motion is more likely to be constant and hence might be more reliable for motion prediction. This is consistent with the psychophysical experiments on motion inertia (Anstis and Ramachandran, 1987) and temporal smoothing of motion signals (Gottsdanker 1956; Snowden and Braddick 1989; Werkhoven et al. 1992) which demonstrated that the direction of motion is the primary cue in temporal motion coherence. Other reasons which might make motion direction more reliable than speed are the unreliability of localspeed measurement due to the poor temporal resolution of local motion units (Kulikowski and Tolhurst, 1973, Holub and MortonGibson 1981) and to involuntary eye movements.
The aim of this paper is to present a new theory for visualmotion estimation and prediction that exploits temporal information. This theory builds on our previous work (Grzywacz et al. 1989; Grzywacz et al. 1995; Yuille et al. 1998) and on recent work on tracking of object boundaries (Isard and Blake 1996). In addition, this paper gives a concrete example of the abstract arguments on temporal grouping proposed by Yuille and Grzywacz (1998). To this end, we use a Bayesian formulation of estimation over time (Ho and Lee 1964), which allows us simultaneously to make predictions over time, update those predictions using new data, and reject data that are inconsistent with the predictions^{2}^{2}2We will be updating the marginal
distributions of velocity at each image point rather than the full probability distribution of the entire velocity field, which would require spatial coupling.
. This rejection of data allows the theory to implement basic temporal grouping (for speculations about more complex temporal grouping, see Yuille and Grzywacz 1998). This basic form of temporal grouping is related to data association as studied by engineers (BarShalom and Fortmann 1988) but differs by being probabilistic (following Ho and Lee 1964). Finally, we show that our temporal grouping theory can be implemented using locally connected networks, which are suggestive of cortical structures.The paper is organized as follows: The theoretical framework for motion prediction and estimation is introduced in Section 2. Section 3 gives a description of the specific model chosen to implement the theory in the continuous time domain. Implementation of the model using a locally connected network is addressed in Section 4 and Section 5. The model’s predictions are tested in Section 6 by performing computer simulations on the motion phenomena described previously. Finally, we discuss relevant issues, such as the possible relationship of the model to biology, in Section 7.
2 The Theoretical Framework
2.1 Background
The theory of stochastic processes gives us a mathematical framework for modelling motion signals that vary over time. These processes can be used to predict the probabilities of future states of the system (e.g. the future velocities) and are called Markov if the probability of future states depends only on their present state (i.e. not
on the time history of how they arrived at this state). In computer vision, Markov processes have been used to model temporal coherence for applications such as the optimal fusing of data from multiple frames of measurements (Matthies
et al. 1989; Clark and Yuille 1990; Chin et al. 1994). Such optimal fusing is typically defined in terms of leastsquares estimates which reduces to Kalman filtering theory (Kalman 1960). Because Kalman filters are (recursive) linear estimators that apply only to Gaussian densities, their applicability in complex scenes involving several moving objects is questionable. In this paper we propose a Bayesian formulation of temporal coherence which is generalization of standard Kalman filters and which can deal with targets moving in complex backgrounds.We begin by deriving a generalization of Kalman filters which is suitable for describing the temporal coherence of simple motion^{3}^{3}3Our development follows the work of Ho and Lee 1964.
. Consider the state vector
describing the status of a system at time (in our theory denotes the velocity field at every point in the image). Our knowledge about how the system might evolve to time is described in a probability distribution function , known as the “prior” (in our theory this prior will encode the temporal coherence assumption). The measurement process that relates the observation of the state vector to its “true” state is described by the likelihood distribution function (in our theory will be the responses of basic motion units at every place in the image). From these two distribution functions, it is straightforward to derive the a posteriori distribution function , which is the distribution of given the whole past set of measurements (note that although the prediction of the future depends only on the current state the estimate of the current state does depend on the entire past history of the measurements). Using Bayes’ rule and a few algebraic manipulations (Ho and Lee 1964), we get(1) 
where is prediction for the future state given the current set of past measurements , and is a normalizing term denoting the confidence in the measure given the set of past measurements . The predictive distribution function can be expressed as
(2) 
Equations 1, 2 are generalizations of the WienerKalman solution for linear estimation in presence of noise. Its evaluation involves two stages. In the prediction stage, given by Equation 2, the probability distribution for the state at time
is determined. This function, which involves the present state and the Markov transition describing the dynamics of the system, has a variance larger than that of
. This increase in the variance results from the nondeterministic nature of the motion. In the measurement stage, performed by Equation 1, the new measurementsare combined using Bayes’ theorem and, if consistent, reinforce the prediction and decrease the uncertainty about the new state. (Inconsistent measurements may increase the uncertainty still further).
It was shown (Ho and Lee 1964) that if the measurement and prior probabilities are both Gaussians, then Equations
1 and 2 reduce to the standard Kalman equations which update the mean and the covariance of andover time. Gaussian distributions, however, are nonrobust (Huber 1981) and an incorrect (outlier) measurement can seriously distort the estimate of the true state. Standard linear Kalman filter models are therefore
not able to account for the psychophysical data which demonstrate, for example, that human observers are able to track targets despite the presence of inconsistent (outlier) measurements (Watamaniuk et al. 1994, Watamaniuk and McKee 1995). Various techniques, known collectively as data association (BarShalom and Fortmann 1988), can be applied to reduce these distortions by using an additional stage which decides whether to reject or accept the new measurements. From the Bayesian perspective, this extra stage is unnecessary and robustness can be ensured by correct choices of the measurement and prior probability models. More specifically, we specify a measurement model which is robust against outliers.2.2 Prediction and Estimation Equations for a Flow Field
We intend to estimate the motion flow field, defined over the twodimensional image array, at each time step. We describe the flow field as , where the parentheses denote variations over each possible spatial position in the image array. The prior model for the motion field and the likelihood for the velocity measurements are described by the probability distribution functions and (see Section 3 for details), where is time increment and represents the response of the local motion measurement units over the image array. Our theory requires that the local velocity measurements are normalized so that for all and at all times (see Section 3 for details). We let be the set of all measurements up to and including time , and be the system’s estimated probability distribution of the velocity field at time . Using Equations 1 and 2 described above, we get
(3) 
for the estimation, and
(4) 
for the prediction.
2.3 The Marginalization Approximation
For simplicity and computational convenience, we assume that the prior and the measurement distributions are chosen to be factorizable in spatial position, so that the probabilities at one spatial point are independent of those at another. This restriction prevents us from including spatial coherence effects which are known to be important aspects of motion perception (see, for example, Yuille and Grzywacz 1988). For the types of stimuli we are considering these spatial effects are of little importance and can be ignored.
In mathematical terms, this spatialfactorization assumption means that we can factorize the prior and the likelihood so that:
(5) 
A further restriction is to modify Equation 4 so that it updates the marginal distributions at each point independently. This again reduces spatial coherence because it decouples the estimates of velocity at each point in space. Once again, we argue that this approximation is valid for the class of stimuli we are considering. This approximation will decrease computation time by allowing us to update the predictions of the velocity distributions at each point independently. The assumption that the measurement model is spatially factorizable means that the estimation stage, Equation 3, can also be performed independently.
This gives update rules for prediction :
(6) 
and for estimation :
(7) 
In what follows we will consider a specific likelihood and prior function which allows us to simplify these equations and define probability distributions on the velocities at each point. This will lead to a formulation for motion prediction and estimation where computation at each spatial position can be performed in parallel.
3 The Model
We now describe a specific model for the likelihood and prior functions which, as we will show, can account for many psychophysical experiments involving temporal motion coherence. Based on this specific prior function, we then derive a formulation for motion prediction in the continuous time domain.
3.1 Likelihood Function
The likelihood function gives a probabilistic interpretation of the measurements given a specific motion. It is probabilistic because the measurement is always corrupted by noise at the input stage. (In many vision theories the likelihood function depends on the image formation process and involves physical constraints such as, for instance, the geometry and surface reflectance. Because we are considering psychophysical stimuli, consisting of moving dots, we can ignore such complications).
To determine our likelihood function, we must first specify the input stage. Ideally, this would involve modeling the behaviors of the cortical cells sensing natural image motion, but this would be too complex to be practical. Instead, we use a simplified model of a bank of receptive fields tuned to various velocities , , and positioned at each image pixel. These cells have observation activities
which are intended to represent the output of a neuronally plausible motion model such as (Grzywacz and Yuille 1990). (In our implementation, these simple model cells receive contributions from the motion of dots falling within a local neighbourhood, typically the four nearest neighbours. Intuitively, the closer the dot is to the center of the receptive field, and the closer its velocity is to the preferred velocity of the cell, then the larger the response. The spatial profile and velocity tuning curve of these receptive fields are described by Gaussian functions whose covariance matrices
and depend on the direction of , and are specified in terms of their longitudinal and transverse components , and , . For more details, see section 4.1 of (Yuille, Burgi, Grzywacz 1997)).The likelihood function specifies the probability of the receptive field responses conditioned on a “true” external motion field. We assume that the measurements depend only on the velocity field at that specific position. We can therefore write . We now specify:
(8) 
where is the joint probability distribution and we set to be^{4}^{4}4This is only one of several possible choices. It is attractive, see (Yuille, Burgi, Grzywacz 1997) because it leads to a simple linear update rule.:
(9) 
with tuning curves given by:
(10) 
where is the covariance matrix which depends on the direction of , and is specified in terms of its longitudinal and transverse components and ). The experimental data (Anstis and Ramachandran 1987, Gottsdanker 1956, Werkhoven et al. 1992) suggest that temporal integration occurs for velocity direction rather than for speed. This is built into our model by choosing the covariances so that the variance is bigger in the direction of motion than in the perpendicular direction (i.e. the velocity component perpendicular to the motion has mean zero and very small variance so the direction of motion is fairly accurate, but the variance of the velocity component along the direction of motion is bigger which means that the estimation of the speed is not accurate). Observe that we have assumed that the response of the measurement device is instantaneous. It would be possible to adapt the likelihood function to allow for a time lag but we have not pursued this. Such a model might be needed to account for motion blurring.
3.2 Prior Function
We now turn to the prior function for position and velocity. For a dot moving at approximately constant velocity, the state transitions for position and velocity is given respectively by , and , where and
are random variables representing uncertainty in the position and velocity. Extending this model to apply to the flow of dots in an image requires a conditional probability distribution
, where the set of spatial positions and the set of allowed velocities are discretized with the spatial positions set to be , and the velocities at to . We are assuming that so that the velocities are predicted independent of each other. We also assume that the prior is built out of two components: (i) the probability that a dot at position with velocity at time will be present at at time , and (ii) the probability that it will have velocity . These components are combined to give:(11) 
where is a normalization factor chosen to ensure that is normalized. The two conditional probability distribution functions in Equation 3.2
are assumed to be normally distributed, and thus,
(12)  
(13) 
These functions express the belief that the dots are, on average, moving along a straight trajectory (Equation 12) with constant velocity (Equation 13). The covariances and , which quantify the statistical deviations from the model, are defined in a coordinate system based on the velocity vector . These matrices are diagonal in this coordinate system and are expressed in terms of their longitudinal and transverse components .
3.3 Continuous Prediction
Computer simulations of Equation 6 (even for the simple case of moving dots) requires largekernel convolutions which require excessive computational time. Instead, we reexpress Equation 6 in the continuous time domain. (Such a reexpression is not always possible and depends on the specific choices of probability distributions. Note that a similar approach has been applied for the computation of stochastic completion fields by Williams and Jacobs 1997a,b). As shown in the Appendix for Gaussian priors, the evolution of for
satisfies a variant of a partial differential equation, known as
Kolmogorov’s forward equation or FokkerPlanck equation. Our equation is a nonstandard variant because, unlike standard Kolmogorov theory, our theory involves probability distributions at every point in space which interact over time. It is given by:(14) 
where is the volume of velocity space , and and are scalar differential operators (for isotropic diffusions, these scalar operators are Laplacian operators). The terms on the right hand side of this equation stand for velocity diffusion (first term), spatial diffusion (second term) with deterministic drift (third term), and normalization (fourth term). This equation is local and describes the spatiotemporal evolution of the probability distributions .
4 The Implementation
We now address the implementation of the update rule for motion estimation (Equations 7). At each spatial position, two separate cells’ populations are assumed, one for coding the likelihood function and another for coding motion estimation. If we were to choose the largekernel convolution for motion prediction (Equation 6), then the network implementation would reduce to two positive linear networks multiplying each other followed by a normalization, illustrated in Figure 1 and see (Yuille et al 1998). But a problem with implementing this network on serial computers is that the range of the connections between motionestimation cells depends on the magnitude of (that is, we would need longrange connections for high speed tunings). Such long range connections occur in biological systems, like the visual cortex, but are not well suited to VLSI or serial computer implementations. Alternatively, using our differential equation for predicting motion (Equation 14) involves only local connections. We now focus on this particular implementation (see Figure 2).
Kolmogorov’s equation can be solved on an arbitrary lattice using a Taylor series expansion. The spatial term, which contains the diffusion and drift terms, can be written so that the partial derivatives at lattice points are functions of the values of the closest neighbor points. For the sake of numerical accuracy, we have chosen a spatial mesh system where the points are arranged hexagonally. Furthermore, we found it convenient to express the velocity vectors in polar coordinates where their norm and angle can be represented on a rectangular mesh. The change of variable is given by . In polar coordinates, the differential operator becomes
(15) 
where and are the variances in the longitudinal and transverse directions, respectively.
We let the activities of the motionestimation cells at time be represented by a vector . The iterative method for determining such activities involves a foursteps scheme. The first three steps are concerned with solving Kolmogorov’s equation. For a cell at position and of velocity tuning , the three steps are:
(16) 
where the coefficients and are functions of , , and the covariance matrices. The constants , represent the quantization factors for and . The superscripts , , indicate the order in which these three steps proceed (i.e. a single time step for the complete system is broken down into four substeps for implementational convenience).
The first step requires evaluating the spatial differential operator and involves six neighbor points (on the hexagonal spatial lattice). The second step calculates the velocity differential operator and involves four neighbor cells. Periodic boundary conditions are assumed in space and velocity. The third step performs the normalization. To guarantee stability of the whole iterative method, the time step
has been determined using von Neumann analysis (see Courant and Hilbert 1953), which considers the independent solutions, or eigenmodes, of finitedifference equations. Typically, the stability conditions involve ratios between the time step and the quantization scales of space and velocity.
If we are performing prediction without measurement, then the fourth step – the determination of the activity of the motionestimation cells – simply reduces to
(17) 
If there are measurements, then we apply the motionestimation equation (Equation 7), which consists of multiplying the motion prediction with the likelihood function (as defined by Equations 8 and 9). Therefore the complete update rule is
(18) 
where is a normalization constant to impose so that we can interpret these activities as the probabilities of the true velocities.
5 Methods
Our model has two extreme forms of behavior depending on how well the input data agree with the model predictions. If the data are generally in agreement with the model s predictions, then it behaves like a standard Kalman ?filter (i.e., with gaussian distributions) and integrates out the noise. At the other extreme, if the data are extremely different from the prediction, then the model treats the data as outliers or distractors and ignores them. We are mainly concerned here with situations where the model rejects data. The precise situation where noise and distractors start getting confused is very interesting, but we do not address it here.
We ?first checked the ability of our model to integrate out weak noisy stimuli for tracking a single dot moving in an empty background. For this situation, temporal grouping (data association) is straightforward, and so standard Kalman ?filters can be used. To evaluate the model, we ran a set of trials that plotted the relative estimation of two velocities as a function of the number of jumps. The graph we obtained (see Figure 3) is very similar to those reported in the psychophysical literature (McKee el al. 1986). Henceforth, we assume that the model is able to integrate out noisy input. For the remaining experiments, we assumed that the local velocity measurements were noisy in the sense that the measurement units specified a probability of velocity measurements rather than a single velocity. However, the probability of velocities itself was not noisy.
We next tested our model on three psychophysical motion phenomena. First, we considered the improved accuracy of velocity estimation of a single dot over time (Snowden and Braddick, 1989; Watamaniuk el al. 1989; Welch et al. 1997; Ascher, 1998). Then we examined the predictions for the motion occlusion experiments (Watamaniuk and McKee, 1995). Finally we investigated the motion outlier experiments (Watamaniuk et al. 1994).
For all simulations we set the parameters as follows. All dots were moving at a speed of 6 jumps per second (jps), where one jump corresponds to the distance separating two neighbouring pixels in the horizontal direction. The longitudinal and transverse components of the covariances matrices were , , , and for the measurement, , and for the likelihood function, and , , , and for the prior function. These parameters were chosen by experimenting with the computer implementation but the results are relatively insensitive to their precise values (i.e. detailed fine tuning was not required).
Except for one of the occluder experiments (the occluder defined by distractor motion), where we used 18 velocity channels (six equidistant directions and three speeds), we used 30 velocity channels positioned at each spatial position, that is, six equidistant directions (thus , starting at the origin), and five speeds (, with slowest speed channel tuned to 2 jps). To guarantee stability of the numerical method we set the time increment to
. Initial conditions were uniformly distributed density functions. There were
spatial positions.6 Results
6.1 Velocity Estimation in Time
For this experiment, the stimulus was a single dot moving in an empty background. To evaluate how the velocity estimation evolves in time, we measured two numbers: The first number was the sharpness of the velocity probability at each position, which we define to be the negative of the entropy of the distribution (the entropy of the distribution is given by ) plus the maximum of the entropy (to ensure that the sharpness is nonnegative). Observe that the sharpness is the KullbackLeibler divergence between the velocity distribution and the uniform distribution (more precisely
.) As is wellknown (Cover and Thomas 1991) the KullbackLeibler divergence is positive semidefinite taking its minimum value of
when and increases the more differs from the uniform distribution (i.e. the sharper the distribution becomes). Hence the higher this sharpness, the more precise the velocity estimate. Moreover, the sharpness will be lowest in positions where there are no dots moving (i.e. for which , ). The target, a coherently moving dot, should have a relatively high sharpness response surrounded by low sharpness responses in neighboring positions. The second number is the confidence factor in Equation 7. This measures how probable the model judges the new input data. The higher this number, the more the new measurement is consistent with the prediction (and hence the more likely that the measurement is due to coherent motion).The results of motion estimation for this experimental stimulus is shown in Figure 4. This figure shows an initial fast decrease in the entropy of the velocity distribution followed by a slow decrease (i.e. an increase in sharpness of the density function). Also shown is the increase in the confidence factor of the target velocity estimation at each new iteration. These two effects indicate the increased accuracy of the velocity estimation with each new iteration, consistent with the psychophysical literature (Snowdon and Braddick 1989, Watamaniuk el al. 1989, Welch et al. 1997, Ascher 1998). Note that for coherent motion the sharpness of the model increases rapidly at first and then appears to slow down. Our results suggest that the model reaches an asymptotic limit although they do not completely rule out a continual gradual increase. We argue that asymptotic behaviour is more likely as both the predictions and the measurements have inherent noise which can never be eliminated. Moreover, it can be proven that the standard (Gaussian) Kalman model will converge to an asymptotic limit depending of the variances of the prior and likelihood functions (and we verified this by computer simulations). It seems also that human observers also reach a similar asymptotic limit and their accuracy does not become infinitely good with increasing number of input stimuli.
We also tested the estimation of velocity for a dot moving on a circular trajectory and our model can successfully estimate the dot’s speed, although the sharpness and confidence are not as high as they are for straight motion (recall that the prior prefers straight trajectories). This again seems consistent with the psychophysics (Watamaniuk et al. 1994, Verghese et al. 1999).
6.2 Single Dot With Occluders
Next, we explored what would happen if the target dot went behind an occluding region where no input measurements could be made (we call this a black occluder). The results were the same as the previous case until the target dot reached the occluder. In the occluded region, the motion model continued to propagate the target dot but, lacking supporting observations, the probability distribution started to diffuse in space and in velocity. This dual diffusion can be seen in Figure 5 where the sharpness of velocity estimation is shown after the dot entered the occluding region. Observe the decrease in sharpness of the density function, indicating a degradation of velocity estimation, when the target is behind the occluder. However, the model still has enough “momentum” to propagate its estimate of the target dot’s position even though no measurements are available (this will, of course, breakdown if the occluder gets too large). This is consistent with the findings by Watamaniuk and McKee (1995) who showed that observers had a higher than average chance of detecting the target when it emerges from the occluder into a noisy background.
We then tested our model with motion occluders, occluders defined by motion flows as described in Watamaniuk and McKee (1995), see Figure 6A. We use the same measures as for the single isolated dot. We also plot the cells’ activities as a function of the direction tuning for the cells tuned to the optimal speed (Figure 6B). The plot indicates that the cells can signal nonzero probabilities for several motions at the same time. After entering the occluding region, the peak corresponding to the target motion gets smaller than the peak due to the occluders. However, the peak of the target remains and so the target peak can rapidly increase as the target exits from the occluders. Our model predicts that is easier to deal with black occluders than with motion occluders which is not consistent with the psychophysics that shows a rough similarity of effects for both types of occluders. We offer two possible explanations for this inconsistency: First, the model’s directionally selective units and/or predictionpropagation mechanism have too broad a directional tuning. Narrowing it may lead to weaker interactions between the motion occluder and the target. Second, in contrast with the black occluders, the motion occluders may group together as a surface, which would help explain why the visual system may be more tolerant of motion occluders. (Such a heightened tolerance might compensate for the initial putative advantage of the black occluders.) Kanizsa (1979) demonstrates several perceptual phenomena, which appear to require this type of explanation. A limitation of our current model is that it does not include such effects.
6.3 Outlier Detection
In the outlier detection experiments (Watamaniuk et al. 1994, Verghese et al. 1999), the target dot is undergoing regular motion but it is surrounded by distractor dots, which are undergoing random Brownian motion. To show the velocity estimation at each position, we plot the response of our network using a rectangle to display the two properties of sharpness and confidence. The width of the rectangle gives the sharpness of the set of cells at that point and the height gives the confidence factor. The arrow shows the mean of the estimated velocity. It can be seen (Figure 7) that the target dot’s signal rapidly reaches large sharpness and confidence by comparison to the distractor dots, which are not moving coherently enough to gain confidence or sharpness. The sharpness of the target dot’s signal does not grow monotonically because the distractor dots sometimes interfere with the target’s motion by causing distracting activity in the measurement cells.
7 Discussion
This work provides a theory for temporal coherence. The theory was formulated in terms of Bayesian estimation using motion measurements and motion prediction. By incorporating robustness in the measurement model, the system could perform temporal grouping (or data association), which enabled it to choose what data should be used to update the state estimation and what data could be ignored. We also derived a continuous form for the prediction equation, and showed that it corresponds to a variant of Kolmogorov’s equations at each node. In deriving our theory, we made an assumption of spatial factorizability of the probability distributions and we made the approximation of updating the marginal distributions of velocity at each point. This allowed us to perform local computations and simplified our implementation. We argued that these assumptions/approximations (as embodied in our choice of prior probability) are suitable for the stimuli we are considering, but that would need to be revised to include spatial coherence effects (Yuille and Grzywacz 1988, Yuille and Grzywacz 1989). The prior distribution used in this paper would need to be modified to incorporate these effects.
In addition, recent results by McKee and Verghese (personal communication) also suggest that that a more sophisticated prior may be needed. Their results seem to show that observers are better at predicting where there is likely to be motion rather than what the velocity will be. Recall that our prior probability model has two predictive components based on the current estimation of velocity at each image position: First, the velocity estimation is used to predict the positions in the image where the velocity cells would be excited. Second, we predict that the new velocity is close to the original velocity. By contrast, McKee and Verghese’s results seem to show that human observers can cope with reversals of motion direction (such as a ball bouncing off a wall). This is a topic of current research and we note that priors which allow for this reversal have already been developed by the computer vision tracking community (Isard and Blake 1998).
Simulations of the model seem in qualitative agreement with a number of existing psychophysical experiments on motion estimation over time, motion occluders and motion outliers (McKee et al. 1986; Watamaniuk and McKee 1995; Watamaniuk et al. 1994). Such a qualitative agreement is characterized by three main features: (i) the final covariance of the motion estimate, (ii) the number of jumps needed to reach such a final covariance, and (iii) the extent of motion inertia during an occlusion, or between two measurements. These three features are controlled as follows: the likelihood’s covariance matrix determines the initial distribution, and thus the number of jumps to reach the desired covariance, while the prior’s covariance matrix affects all three features by imposing an upper bound to the covariance and by setting the time constants of the diffusion process between measurements.
Implementation of the model’s equations led to a network that shares interesting similarities with known properties of the visual cortex. It involves a columnar structure where the columns, at each spatial point, are composed of a set of cells tuned to a range of velocities (columnar structures exist in the cortex but none have yet been found for velocities). The observationcell layer involves local connections within the columns to compute the likelihood function. The estimated velocity flow is computed in a second layer. If these layers exist, it would be natural for them to lie in cortical areas involved with motion such as MT and MST (Maunsell and Newsome 1987, Merigan and Maunsell 1993). The estimation layer carries out calculation according to Kolmogorov’s equation, or the discrete long range version described in (Yuille et al 1998), and consists of spatial columns of cells which excite locally each other to predict the expected velocities in the future. The calculation requires excitatory and inhibitory synapses (i.e. no “negative synapses”). The inputs from the observation layer multiply the predictions in the estimated layer. Finally, we postulate that inhibition occurs in each column to ensure normalization of the velocity estimates at each column. This normalization could be implemented by lateral inhibition as proposed by Heeger (1992).
A difficult aspect of this network, as a biological model, is its need for multiplications (such multiplications arise naturally from our probabilistic formulation using the rules for combining probabilities). Neuronal multiplication mechanisms were argued for by Reichardt (1961), Poggio and Reichardt (1973, 1976) on theoretical grounds. A specific biophysical model to approximate multiplication was proposed by Torre and Poggio (1978). Detailed investigations of this model, however, showed that it provided at best a rough approximation for multiplication (Grzywacz and Koch 1987). Moreover, experiments showed that motion computations in the rabbit retina, for which the model was originally proposed, were not well approximated by multiplications (Grzywacz et al. 1990). Recent related work (Mel et al. 1988, Mel 1993), though arguing for multiplicationlike processing, has not attempted to attain good approximations to multiplication. On the other hand, the complexities of neurons make it hard to rule out the existence of domains in which they perform multiplications. Overall, it seems best to be agnostic about neuronal multiplication and trust that more detailed experiments will resolve this issue. More pragmatically, it may be best to favour neuronal models which correspond to known biophysical mechanisms. The biophysics of neurons may prevent them from performing “perfect” Bayesian computations and perhaps one should instead seek optimal models which respect these biophysical limitations.
Finally, we emphasize that simple networks of the type we propose can implement Bayesian generalizations of Kalman filters for estimation and prediction. The popularity of standard linear Kalman filters is often due to pragmatic reasons of computational efficiency. Thus linear Kalman filters are often used even when they are inappropriate as models of the underlying processes. In contrast, the main drawback of such Bayesian generalizations is the high computational cost, even though statistical sampling methods can be used successfully in some cases (see Isard and Blake 1996). Our study shows that these computational costs may be better dealt with by using networks with local connections and we are investigating the possibility of implementing our model on parallel architectures in the expectation that we will then be able to do Bayesian estimation of motion in real time.
Appendix
In this Appendix, we derive Kolmogorov’s equation for motion prediction. Let consider Equation 6 for : Taking this limit is tricky and requires Itô calculus (Jazwinski 1970). Our case is also nonstandard because our theory involves probability distributions at every point in space, which interact over time. This means that we cannot simply use the standard Kolmogorov’s equations, which apply to a single distribution only. Here we give a simple derivation, however, which uses delta functions and is appropriate when the prior model is based on a Gaussian . The prior specified by equations 3.2, 12 is written as:
(A1) 
The normalization term in the denominator is used to ensure that the probabilities of the velocities integrate to one at each spatial point . In this equation, we have scaled the covariances by . This is necessary for taking the limit as (Jazwinski 1970).
Between observations we can express the evolution of the prior density as:
(A2) 
We now perform a Taylor series expansion of in powers of keeping the zeroth and first order terms (higher order terms will vanish when we take the limit as ). To perform this expansion, we use the assumption that this distribution is expressed in terms of Gaussians. As these Gaussians will tend towards delta functions and the derivatives of Gaussians will tend towards derivatives of delta functions (thereby simplifying the expansion). This derivation can be justified rigourously, playing detailed attention to convergence issues, by the use of distribution theory or the application of Itô calculus. If we expand about we find that the zeroth order term is a Dirac delta function with argument . This term can therefore be integrated out. The derivative with respect to will effectively correspond to differentiating the Gaussian with respect to the covariance. By standard properties of the Gaussian this will be equivalent to a secondorder spatial derivative. We will get similar terms if we expand , but we will also get an additional drift term from differentiating the argument. In addition, we will get other terms from the denominator of Equation A1. (These are required to normalize the distributions and are nonstandard. They are needed because we have a differential equation for a set of interacting probability distributions while the standard Kolmogorov’s equation is for a single probability distribution.) We collect all these zeroth and first order terms in the expansion and substitute into Equation A2. We can then evaluate the integrals using known properties of the delta functions. After some algebra, these integrals yield our variant of Kolmogorov’s equation:
(A3) 
where and are scalar differential operators, and is the volume of velocity space .
Acknowledgments
This work was supported by the AFOSR grant F496209510265 to A.L.Y. and N.M.G. Further support to A.L.Y. came from ARPA and the Office of Naval Research with grant number N000149511022. In addition, N.M.G. received support from grants by the National Eye Institute (EY08921 and EY11170) and the William A. Kettlewell chair. Finally, we thank the National Eye Institute for a core grant to SmithKettlewell (EY06883).
References

Adelson, E.H., and Bergen, J.R. 1985. Spatiotemporal energy models for the perception of motion. J. Opt. Soc. A., 2, 284299.

Anstis, S.M., and Ramachandran, V.S. 1987. Visual inertia in apparent motion. Vision Res. 27, 755764.

Ascher, D. 1998. ”Human Visual Speed Perception: Psychophysics and Modeling.” Program of Cognitive Science. Department of Cognitive and Linguistic Sciences. Brown University.

BarShalom, Y., and Fortmann, T.E. 1988. Tracking and Data Association. Academic Press.

Burr, D.C., Ross, J., and Morrone, M.C. 1986. Seeing objects in motion. Proc. Royal Soc. Lond. B, 227, 249265.

Chin, T.M., Karl, W.C., and Willsky, A.S. 1994. Probabilistic and sequential computation of optical flow using temporal coherence. IEEE Trans. Image Proc. 3, 773788.

Clark, J.J., and Yuille, A.L. 1990. Data Fusion for Sensory Information Processing Systems. Kluwer Academic Press, Boston.

Courant, R., and Hilbert, D. 1953. Methods of Mathematical Physics, Volume 1. Interscience, New York.

Cover, T.M. and J.A. Thomas, J.A. Elements of Information Theory. Wiley Interscience Press. New York. 1991.

Gottsdanker, R.M. 1956. The ability of human operators to detect acceleration of target motion. Psych. Bull. 53, 477487.

Grzywacz, N.M., Amthor, F.R. and Mistler, L.A. 1990. Applicability of quadratic and threshold models to motion discrimination in the rabbit retina. Biol. Cybern. 64, 4149.

Grzywacz, N.M. and Koch, C. 1987. Functional properties of models for direction selectivity in the retina. Synapse, 1, 417434.

Grzywacz, N.M., Smith, J.A., and Yuille, A.L. 1989. A computational framework for visual motion’s spatial and temporal coherence. Proc. IEEE Workshop on Visual Motion, Irvine.

Grzywacz, N.M., Watamaniuk, S.N.J., and McKee, S.P. 1995. Temporal coherence theory for the detection and measurement of visual motion. Vision Res. 35, 31833203.

Grzywacz, N.M. and Yuille, A.L. 1990. “A model for the estimate of local image velocity by cells in the visual cortex”. Proc. Royal Soc. Lond. B, 239, 129161.

Hassenstein, B., and Reichardt, W.E. 1956. Systemtheoretische analyse der zeit, reihenfolgen und vorzeichenauswertung bei der bewegungsperzeption des russelkafers Chlorophanus. Z. Naturforsch. 11b, 513524.

Heeger, D.J. 1992. Normalization of cell responses in cat striate cortex. Visual Neurosci. 9, 181197.

Ho, YC., and Lee, R.C.K. 1964. A Bayesian approach to problems in stochastic estimation and control. IEEE Trans. on Automatic Control 9, 333339.

Holub, R.A. and MortonGibson, M. 1981. Response of visual cortical neurons of the cat to moving sinusoidal gratings: Responsecontrast functions and spatiotemporal integration J. Neurophysiol. 46, 1244–1259.

Huber, P.J. 1981. Robust Statistics. John Wiley and Sons, New York.

Isard, M., and Blake A. 1996. Contour tracking by stochastic propagation of conditional density. Proc. Europ. Conf. Comput. Vision, pp. 343356, Cambridge, UK.

Isard, M., and Blake, A. 1998. A mixedstate condensation tracker with automatic modelswitching. Proc. Int. Conf. Comp. Vis., pp. 107 112, Mumbai, India.

Jazwinski, A.H. 1970. Stochastic Processes and Filtering Theory, Academic Press, San Diego.

Kalman, R.E. 1960. A new approach to linear filtering and prediction problems. Trans. ASME, D: J. Basic Engineering, 82, 3545.

Koch, C., and Poggio, T. 1992. Multiplying with synapses and neurons. In Single Neuron Computation, T. McKenna, J. Davis, and S.F. Zornetzer eds, pp. 315–345. Academic Press, Boston.

Kulikowski, J. J. and Tolhurst, D. J. 1973. Psychophysical evidence for sustained and transient detectors in human vision. Journal of Physiology, 232, 519548.

Matthies, L., Kanade, T., and Szeliski, R. 1989. Kalman filterbased algorithms for estimating depth from image sequences. Int. J. Comput. Vision 3, 209236.

Maunsell, J. H. R. and Newsome, W. T. 1987. Visual Processing in Monkey Extrastriate Cortex. In Cowan, W. M., Shooter, E. M., Stevens, C. F. and Thompson, R. F. (Eds.) Annual Reviews of Neuroscience (vol. 10, pp. 363–401). Palo Alto, CA: Annual Reviews Inc.

Mel B.W., Ruderman D.L., Archie K.A. 1988. Translationinvariant orientation tuning in visual ”complex” cells could derive from intradendritic computations. J Neurosci., 18, 43254334

Mel B.W. 1993. Synaptic integration in an excitable dendritic tree. J. Neurophysiol 70, 10861101.

McKee, S.P., and Welch, L. 1985. Sequential recruitment in the discrimination of velocity. J. Optical Soc. America A 2, 243251.

McKee, S.P., Silverman, G.H., and Nakayama, K. 1986. Precise velocity discrimination despite random variations in temporal frequency and contrast. Vision Res. 26, 609619.

Merigan WH, Maunsell JH. 1993. How parallel are the primate visual pathways? In Cowan, W. M., Shooter, E. M., Stevens, C. F. & Thompson, R. F. (Eds.) Annual Reviews of Neuroscience (vol. 16, pp. 369–4021). Palo Alto, CA: Annual Reviews Inc.

Nakayama, K. 1985. Biological image motion processing: A review. Vision Res. 25, 625660.

Poggio T and Reichardt WE. 1973. Considerations on models of movement detection. Kybernetik, 13, 223227.

Poggio, T. and Reichardt, W.E. 1976. Visual control of orientation behaviour in the fly: Part II: Towards the underlying neural interactions. Q. Rev. Biophys., 9, 377438.

Ramachandran, V.S., and Anstis, S.M. 1983. Extrapolation of motion path in human visual perception. Vision Res. 23, 8385.

Reichardt, W. 1961. Autocorrelation, a principle for the evaluation of sensory information by the nervous system. In Sensory Communication Ed Rosenblith, Wiley, New York.

Snowden, R.J., and Braddick, O.J. 1989. Extension of displacement limits in multipleexposure sequences of apparent motion. Vision Res. 29, 17771787.

van Santen, J.P.H., and Sperling, G. 1984. A temporal covariance model of motion perception J. Opt. Soc. A., A 1, 451473.

Torre, V and Poggio, T. 1978. A synaptic mechanism possibly underlying directional selectivity to motion. Proc. R. Soc. Lond. B., 202, 409416.

Verghese, P., Watamaniuk, S.N.J., McKee, S.P. and Grzywacz, N.M. 1999. Local motion detectors cannot account for the detectability of an extended motion trajectory in noise. Vision Res. 39, 1930.

Watamaniuk, S. N. J., Sekuler, R. & Williams, D. W. 1989. Direction perception in complex dynamic displays: the integration of direction information. Vision Res., 29, 4759.

Watamaniuk, S.N.J., McKee, S.P., and Grzywacz, N.M. 1994. Detecting a Trajectory Embedded in RandomDirection Motion Noise. Vision Res. 35, 6577.

Watamaniuk, S.N.J., and McKee, S.P. 1995. Seeing motion behind occluders. Nature 377, 729730.

Watson, A.B., and Ahumada, A.J. 1985. Model of human visual motion sensing. J. Opt. Soc. A, A 2, 322342.

Welch L., MacLeod D.I., Mckee S.P. 1997 Motion interference: perturbing perceived direction. Vision Res., 37, 27252736.

Werkhoven, P., Snippe, H.P., and Toet, A. 1992. Visual processing of optic acceleration. Vision Res. 32, 23132329.

Williams, L.R., and Jacobs, D.W. 1997a. Local parallel computation of stochastic completion fields. Neural Comput. 9, 859881.

Williams, L.R., and Jacobs, D.W. 1997b. Stochastic completion fields: a neural model of illusory contour shape and salience. Neural Comput. 9, 837858.

Yuille, A.L., Burgi, P.Y., and Grzywacz, N.M. 1998. Visual motion estimation and prediction: A probabilistic network model for temporal coherence. SmithKettlewell Eye Research Institute Technical Report.

Yuille, A.L., Burgi, P.Y., and Grzywacz, N.M. 1998. Visual motion estimation and prediction: A probabilistic network model for temporal coherence. In Proceedings of the Sixth International Conference on Computer Vision, (pp. 973978). IEEE Computer Society.

Yuille, A.L. and Grzywacz, N.M. 1988. “A Computational Theory for the Perception of Coherent Visual Motion”. Nature. 333, pp 7174.

Yuille, A.L. and Grzywacz, N.M. 1989. A Mathematical Analysis of the Motion Coherence Theory. International Journal of Computer Vision, 3, 155175.

Yuille, A.L., and Grzywacz, N.M. 1998. A Theoretical framework for visual motion. In HighLevel Motion Processing : Computational, Neurophysiological, and Psychophysical Perspectives, T. Watanabe ed. MIT Press.

Yuille, A.L., and Grzywacz, N.M. A computational theory for the perception of coherent visual motion. Nature. 333. pp 7174. 1988.
Comments
There are no comments yet.