[AI Robotics Summer School 2016] From local features to deep learning
We introduce a novel Deep Network architecture that implements the full feature point handling pipeline, that is, detection, orientation estimation, and feature description. While previous works have successfully tackled each one of these problems individually, we show how to learn to do all three in a unified manner while preserving end-to-end differentiability. We then demonstrate that our Deep pipeline outperforms state-of-the-art methods on a number of benchmark datasets, without the need of retraining.READ FULL TEXT VIEW PDF
[AI Robotics Summer School 2016] From local features to deep learning
Code release for the ECCV 2016 paper
Code release for the ECCV 2016 paper
Local features play a key role in many Computer Vision applications. Finding and matching them across images has been the subject of vast amounts of research. Until recently, the best techniques relied on carefully hand-crafted features [1, 2, 3, 4, 5]
. Over the past few years, as in many areas of Computer Vision, methods based in Machine Learning, and more specifically Deep Learning, have started to outperform these traditional methods[6, 7, 8, 9, 10].
These new algorithms, however, address only a single step in the complete processing chain, which includes detecting the features, computing their orientation, and extracting robust representations that allow us to match them across images. In this paper we introduce a novel Deep architecture that performs all three steps together. We demonstrate that it achieves better overall performance than the state-of-the-art methods, in large part because it allows these individual steps to be optimized to perform well in conjunction with each other.
Our architecture, which we refer to as LIFT for Learned Invariant Feature Transform, is depicted by Fig. 1
. It consists of three components that feed into each other: the Detector, the Orientation Estimator, and the Descriptor. Each one is based on Convolutional Neural Networks (CNNs), and patterned after recent ones[6, 9, 10] that have been shown to perform these individual functions well. To mesh them together we use Spatial Transformers  to rectify the image patches given the output of the Detector and the Orientation Estimator. We also replace the traditional approaches to non-local maximum suppression (NMS) by the soft argmax function . This allows us to preserve end-to-end differentiability, and results in a full network that can still be trained with back-propagation, which is not the case of any other architecture we know of.
Also, we show how to learn such a pipeline in an effective manner. To this end, we build a Siamese network and train it using the feature points produced by a Structure-from-Motion (SfM) algorithm that we ran on images of a scene captured under different viewpoints and lighting conditions, to learn its weights. We formulate this training problem on image patches extracted at different scales to make the optimization tractable. In practice, we found it impossible to train the full architecture from scratch, because the individual components try to optimize for different objectives. Instead, we introduce a problem-specific learning approach to overcome this problem. It involves training the Descriptor first, which is then used to train the Orientation Estimator, and finally the Detector, based on the already learned Descriptor and Orientation Estimator, differentiating through the entire network. At test time, we decouple the Detector, which runs over the whole image in scale space, from the Orientation Estimator and Descriptor, which process only the keypoints.
In the next section we briefly discuss earlier approaches. We then present our approach in detail and show that it outperforms many state-of-the-art methods.
The amount of literature relating to local features is immense, but it always revolves about finding feature points, computing their orientation, and matching them. In this section, we will therefore discuss these three elements separately.
Research on feature point detection has focused mostly on finding distinctive locations whose scale and rotation can be reliably estimated. Early works [13, 14] used first-order approximations of the image signal to find corner points in images. FAST  used Machine Learning techniques but only to speed up the process of finding corners. Other than corner points, SIFT  detect blobs in scale-space; SURF  use Haar filters to speed up the process; Maximally Stable Extremal Regions (MSER)  detect regions;  detect affine regions. SFOP  use junctions and blobs, and Edge Foci  use edges for robustness to illumination changes. More recently, feature points based on more sophisticated and carefully designed filter responses [5, 20] have also been proposed to further enhance the performance of feature point detectors.
In contrast to these approaches that focus on better engineering, and following the early attempts in learning detectors [21, 22],  showed that a detector could be learned to deliver significantly better performance than the state-of-the-art. In this work, piecewise-linear convolutional filters are learned to robustly detect feature points in spite of lighting and seasonal changes. Unfortunately, this was done only for a single scale and from a dataset without viewpoint changes. We therefore took our inspiration from it but had to extend it substantially to incorporate it into our pipeline.
Despite the fact that it plays a critical role in matching feature points, the problem of estimating a discriminative orientation has received noticeably less attention than detection or feature description. As a result, the method introduced by SIFT  remains the de facto standard up to small improvements, such as the fact that it can be sped-up by using the intensity centroid, as in ORB .
A departure from this can be found in a recent paper  that introduced a Deep Learning-based approach to predicting stable orientations. This resulted in significant gains over the state-of-the-art. We incorporate this architecture into our pipeline and show how to train it using our problem-specific training strategy, given our learned descriptors.
Feature descriptors are designed to provide discriminative representations of salient image patches, while being robust to transformations such as viewpoint or illumination changes. The field reached maturity with the introduction of SIFT , which is computed from local histograms of gradient orientations, and SURF , which uses integral image representations to speed up the computation. Along similar lines, DAISY  relies on convolved maps of oriented gradients to approximate the histograms, which yields large computational gains when extracting dense descriptors.
Even though they have been extremely successful, these hand-crafted descriptors can now be outperformed by newer ones that have been learned. These range from unsupervised hashing to supervised learning techniques based on linear discriminant analysis[23, 24]25], and convex optimization . An even more recent trend is to extract features directly from raw image patches with CNNs trained on large volumes of data. For example, MatchNet  trained a Siamese CNN for feature representation, followed by a fully-connected network to learn the comparison metric. DeepCompare  showed that a network that focuses on the center of the image can increase performance. The approach of  relied on a similar architecture to obtain state-of-the-art results for narrow-baseline stereo. In , hard negative mining was used to learn compact descriptors that use on the Euclidean distance to measure similarity. The algorithm of  relied on sample triplets to mine hard negatives.
In this work, we rely on the architecture of  because the corresponding descriptors are trained and compared with the Euclidean distance, which has a wider range of applicability than descriptors that require a learned metric.
In this section, we first formulate the entire feature detection and description pipeline in terms of the Siamese architecture depicted by Fig. 2. Next, we discuss the type of data we need to train our networks and how to collect it. We then describe the training procedure in detail.
We use image patches as input, rather than full images. This makes the learning scalable without loss of information, as most image regions do not contain keypoints. The patches are extracted from the keypoints used by a SfM pipeline, as will be discussed in Section 3.2. We take them to be small enough that we can assume they contain only one dominant local feature at the given scale, which reduces the learning process to finding the most distinctive point in the patch.
To train our network we create the four-branch Siamese architecture pictured in Fig. 2. Each branch contains three distinct CNNs, a Detector, an Orientation Estimator, and a Descriptor. For training purposes, we use quadruplets of image patches. Each one includes two image patches and , that correspond to different views of the same 3D point, one image patch , that contains the projection of a different 3D point, and one image patch that does not contain any distinctive feature point. During training, the -th patch of each quadruplet will go through the -th branch.
To achieve end-to-end differentiability, the components of each branch are connected as follows:
Given an input image patch , the Detector provides a score map .
We perform a soft argmax  on the score map and return the location of a single potential feature point.
We extract a smaller patch centered on with the Spatial Transformer layer Crop (Fig. 2). This serves as the input to the Orientation Estimator.
The Orientation Estimator predicts a patch orientation .
We rotate according to this orientation using a second Spatial Transformer layer, labeled as Rot in Fig. 2, to produce .
is fed to the Descriptor network, which computes a feature vector .
Note that the Spatial Transformer layers are used only to manipulate the image patches while preserving differentiability. They are not learned modules. Also, both the location proposed by the Detector and the orientation for the patch proposal are treated implicitly, meaning that we let the entire network discover distinctive locations and stable orientations while learning.
Since our network consists of components with different purposes, learning the weights is non-trivial. Our early attempts at training the network as a whole from scratch were unsuccessful. We therefore designed a problem-specific learning approach that involves learning first the Descriptor, then the Orientation Estimator given the learned descriptor, and finally the Detector, conditioned on the other two. This allows us to tune the Orientation Estimator for the Descriptor, and the Detector for the other two components.
There are datasets that can be used to train feature descriptors  and orientation estimators . However it is not so clear how to train a keypoint detector, and the vast majority of techniques still rely on hand-crafted features. The TILDE detector  is an exception, but the training dataset does not exhibit any viewpoint changes.
To achieve invariance we need images that capture views of the same scene under different illumination conditions and seen from different perspectives. We thus turned to photo-tourism image sets. We used the collections from Piccadilly Circus in London and the Roman Forum in Rome from  to reconstruct the 3D using VisualSFM , which relies of SIFT features. Piccadilly contains 3384 images, and the reconstruction has 59k unique points with an average of 6.5 observations for each. Roman-Forum contains 1658 images and 51k unique points, with an average of 5.2 observations for each. Fig. 3 shows some examples.
We split the data into training and validation sets, discarding views of training points on the validation set and vice-versa. To build the positive training samples we consider only the feature points that survive the SfM reconstruction process. To extract patches that do not contain any distinctive feature point, as required by our training method, we randomly sample image regions that contain no SIFT features, including those that were not used by SfM.
We extract grayscale training patches according to the scale of the point, for both feature and non-feature point image regions. Patches are extracted from a support region at these locations, and standardized into pixels where . The smaller patches and that serve as input to the Orientation Estimator and the Descriptor, are cropped and rotated versions of these patches, each having size , where . The smaller patches effectively correspond to the SIFT descriptor support region size of . To avoid biasing the data, we apply uniform random perturbations to the patch location with a range of (4.8
). Finally, we normalize the patches with the grayscale mean and standard deviation of the entire training set.
Learning feature descriptors from raw image patches has been extensively researched during the past year [7, 8, 10, 27, 28, 31], with multiple works reporting impressive results on patch retrieval, narrow baseline stereo, and matching non-rigid deformations. Here we rely on the relatively simple networks of , with three convolutional layers followed by hyperbolic tangent units, pooling  and local subtractive normalization, as they do not require learning a metric.
The Descriptor can be formalized simply as
where denotes the Descriptor CNN, its parameters, and is the rotated patch from the Orientation Estimator. When training the Descriptor, we do not yet have the Detector and the Orientation Estimator trained. We therefore use the image locations and orientations of the feature points used by the SfM to generate image patches .
We train the Descriptor by minimizing the sum of the loss for pairs of corresponding patches and the loss for pairs of non-corresponding patches . The loss for pair is defined as the hinge embedding loss of the Euclidean distance between their description vectors. We write
where positive and negative samples are pairs of patches that do or do not correspond to the same physical 3D points, is the Euclidean distance, and is the margin for embedding.
We use hard mining during training, which was shown in  to be critical for descriptor performance. Following this methodology, we forward sample pairs and use only the pairs with the highest training loss for back-propagation, where is the ‘mining ratio’. In  the network was pre-trained without mining and then fine-tuned with . Here, we use an increasing mining scheme where we start with and double the mining ratio every 5000 batches. We use balanced batches with 128 positive pairs and 128 negative pairs, mining each separately.
Our Orientation Estimator is inspired by that of . However, this specific one requires pre-computations of description vectors for multiple orientations to compute numerically the Jacobian of the method parameters with respect to orientations. This is a critical limitation for us because we treat the output of the detector component implicitly throughout the pipeline and it is thus not possible to pre-compute the description vectors.
We therefore propose to use Spatial Transformers  instead to learn the orientations. Given a patch from the region proposed by the detector, the Orientation Estimator predicts an orientation
where denotes the Orientation Estimator CNN, and its parameters.
Together with the location from the Detector and the original image patch, is then used by the second Spatial Transformer Layer to provide a patch , which is the rotated version of patch .
We train the Orientation Estimator to provide the orientations that minimize the distances between description vectors for different views of the same 3D points. We use the already trained Descriptor to compute the description vectors, and as the Detector is still not trained, we use the image locations from SfM. More formally, we minimize the loss for pairs of corresponding patches, defined as the Euclidean distance between their description vectors
where is the patch centered on after orientation correction: . This complex notation is necessary to properly handle the cropping of the image patches. Recall that pairs comprise image patches containing the projections of the same 3D point, and locations and denote the reprojections of these 3D points. As in , we do not use pairs that correspond to different physical points whose orientations are not related.
The Detector takes an image patch as input, and returns a score map. We implement it as a convolution layer followed by piecewise linear activation functions, as in TILDE. More precisely, the score map for patch is computed as:
where denotes the Detector itself with parameters , is if
is odd andotherwise, is made of the filters and biases of the convolution layer to learn, denotes the convolution operation, and and are hyper-parameters controlling the complexity of the piecewise linear activation function.
The main difference with TILDE lies in the way we train this layer. To let have maxima in places other than a fixed location retrieved by SfM, we treat this location implicitly, as a latent variable. Our method can potentially discover points that are more reliable and easier to learn, whereas  cannot. Incidentally, in our early experiments, we noticed that it was harmful to force the Detector to optimize directly for SfM locations.
From the score map , we obtain the location of a feature point as
where softargmax is a function which computes the Center of Mass with the weights being the output of a standard softmax function . We write
where are locations in , and is a hyper-parameter controlling the smoothness of the softargmax. This softargmax function acts as a differentiable version of non-maximum suppression. is given to the first Spatial Transformer Layer together with the patch to extract a smaller patch used as input to the Orientation Estimator.
As the Orientation Estimator and the Descriptor have been learned by this point, we can train the Detector given the full pipeline. To optimize over the parameters , we minimize the distances between description vectors for the pairs of patches that correspond to the same physical points, while maximizing the classification score for patches not corresponding to the same physical points.
More exactly, given training quadruplets , where and correspond to the same physical point, and correspond to different SfM points, and
to a non-feature point location, we minimize the sum of their loss functions
where is a hyper-parameter balancing the two terms in this summation
with and if , and and otherwise to balance the positives and negatives. softmax is the log-mean-exponential softmax function. We write
Note that the locations of the detected feature points appear only implicitly and are discovered during training. Furthermore, all three components are tied in with the Detector learning. As with the Descriptor we use a hard mining strategy, in this case with a fixed mining ratio of .
In practice, as the Descriptor already learns some invariance, it can be hard for the Detector to find new points to learn implicitly. To let the Detector start with an idea of the regions it should find, we first constrain the patch proposals that correspond to the same physical points to overlap. We then continue training the Detector without this constraint.
Specifically, when pre-training the Detector, we replace in Eq. (8) with , where is equal to 0 when the patch proposals overlap exactly, and increases with the distance between them otherwise. We therefore write
where , , is the norm. Recall that pixels is the width and height of the patch proposals.
The pipeline used at run-time is shown in Fig. 4. As our method is trained on patches, simply applying it over the image would require the network to be tested with a sliding window scheme over the whole image. In practice, this would be too expensive. Fortunately, as the Orientation Estimator and the Descriptor only need to be run at local maxima, we can simply decouple the detector from the rest to apply it to the full image, and replace the softargmax function by NMS, as outlined in red in Fig. 4. We then apply the Orientation Estimator and the Descriptor only to the patches centered on local maxima.
More exactly, we apply the Detector independently to the image at different resolutions to obtain score maps in scale space. We then apply a traditional NMS scheme similar to that of  to detect feature point locations.
In this section, we first present the datasets and metrics we used. We then present qualitative results, followed by a thorough quantitative comparison against a number of state-of-the-art baselines, which we consistently outperform.
Finally, to better understand what elements of our approach most contribute to this result, we study the importance of the pre-training of the Detector component, discussed in Section 3.5, and analyze the performance gains attributable to each component.
We evaluate our pipeline on three standard datasets:
The Strecha dataset , which contains 19 images of two scenes seen from increasingly different viewpoints.
The DTU dataset , which contains 60 sequences of objects with different viewpoints and illumination settings. We use this dataset to evaluate our method under viewpoint changes.
The Webcam dataset , which contains 710 images of 6 scenes with strong illumination changes but seen from the same viewpoint. We use this dataset to evaluate our method under natural illumination changes.
For Strecha and DTU we use the provided ground truth to establish correspondences across viewpoints. We use a maximum of 1000 keypoints per image, and follow the standard evaluation protocol of  on the common viewpoint region. This lets us evaluate the following metrics.
Repeatability (Rep.): Repeatability of feature points, expressed as a ratio. This metric captures the performance of the feature point detector by reporting the ratio of keypoints that are found consistently in the shared region.
Nearest Neighbor mean Average Precision (NN mAP): Area Under Curve (AUC) of the Precision-Recall curve, using the Nearest Neighbor matching strategy. This metric captures how discriminating the descriptor is by evaluating it at multiple descriptor distance thresholds.
Matching Score (M. Score): The ratio of ground truth correspondences that can be recovered by the whole pipeline over the number of features proposed by the pipeline in the shared viewpoint region. This metric measures the overall performance of the pipeline.
We compare our method on the three datasets to the following combination of feature point detectors and descriptors, as reported by the authors of the corresponding papers: SIFT , SURF , KAZE , ORB , Daisy  with SIFT detector, sGLOH  with Harris-affine detector , MROGH  with Harris-affine detector, LIOP  with Harris-affine detector, BiCE  with Edge Foci detector , BRISK , FREAK  with BRISK detector, VGG  with SIFT detector, DeepDesc  with SIFT detector, PN-Net  with SIFT detector, and MatchNet  with SIFT detector. We also consider SIFT with Hessian-Affine keypoints . For the learned descriptors VGG, DeepDesc, PN-Net and MatchNet we use SIFT keypoints because they are trained using a dataset created with Difference-of-Gaussians, which is essentially the same as SIFT. In the case of Daisy, which was not developed for a specific detector, we also use SIFT keypoints. To make our results reproducible, we provide additional implementation details for LIFT and the baselines in the supplementary material.222Source and models will be available at https://github.com/cvlab-epfl/LIFT.
Fig. 5 shows image matching results with 500 feature points, for both SIFT and our LIFT pipeline trained with Piccadilly. As expected, LIFT returns more correct correspondences across the two images. One thing to note is that the two DTU scenes in the bottom two rows are completely different from the photo-tourism datasets we used for training. Given that the two datasets are very different, this shows good generalization properties.
Fig. 6 shows the average matching score for all three datasets, and Table 1 provides the exact numbers for the two LIFT variants. LIFT (pic) is trained with Piccadilly and LIFT (rf) with Roman-Forum. Both of our learned models significantly outperform the state-of-the-art on Strecha and DTU and achieve state-of-the-art on Webcam. Note that KAZE, which is the best performing competitor on Webcam, performs poorly on the other two datasets. As discussed above, Piccadilly and Roman-Forum are very different from the datasets used for testing. This underlines the strong generalization capability of our approach, which is not always in evidence with learning-based methods.
Interestingly, on DTU, SIFT is still the best performing method among the competitors, even compared to methods that rely on Deep Learning, such as DeepDesc and PN-Net. Also, the gap between SIFT and the learning-based VGG, DeepDesc, and PN-Net is not large for the Strecha dataset.
These results show that although a component may outperform another method when evaluated individually, they may fail to deliver their full potential when integrated into the full pipeline, which is what really matters. In other words, it is important to learn the components together, as we do, and to consider the whole pipeline when evaluating feature point detectors and descriptors.
|Trained on Piccadilly||Trained on Roman-Forum|
|Trained on Piccadilly||Trained on Roman-Forum|
|Det.||Ori.||Desc.||Rep.||NN mAP||M.Score||Rep.||NN mAP||M.Score|
Recall that we pre-train the detector and then finalize the training with the Orientation Estimator and the Descriptor, as discussed in Section 3.5. It is therefore interesting to see the effect of this finalizing stage. In Table 2 we evaluate the entire pipeline with the pre-trained Detector and the final Detector. As the pair-wise loss term of Eq. (11) is designed to emulate the behavior of an ideal descriptor, the pre-trained Detector already performs well. However, the full training pushes the performance slightly higher.
A closer look at Table 2 reveals that gains are larger overall for Piccadilly than for Roman-Forum
. This is probably due to the fact thatRoman-Forum does not have many non-feature point regions. In fact, the network started to over-fit quickly after a few iterations on this dataset. The same happened when we further tried to fine-tune the full pipeline as a whole, suggesting that our learning strategy is already providing a good global solution.
To understand the influence of each component on the overall performance, we exchange them with their SIFT counterparts, for both LIFT (pic) and LIFT (rf), on Strecha. We report the results in Table 3. In short, each time we exchange to SIFT, we decrease performance, thus showing that each element of the pipeline plays and important role. Our Detector gives higher repeatability for both models. Having better orientations also helps whichever detector or descriptor is being used, and also the Deep Descriptors perform better than SIFT.
One thing to note is that our Detector is not only better in terms of repeatability, but generally better in terms of both the NN mAP, which captures the descriptor performance, and in terms of matching score, which evaluates the full pipeline. This shows that our Detector learns to find not only points that can be found often but also points that can be matched easily, indicating that training the pipeline as a whole is important for optimal performance.
We have introduced a novel Deep Network architecture that combines the three components of standard pipelines for local feature detection and description into a single differentiable network. We used Spatial Transformers together with the softargmax function to mesh them together into a unified network that can be trained end-to-end with back-propagation. While this makes learning the network from scratch theoretically possible, it is not practical. We therefore proposed an effective strategy to train it.
Our experimental results demonstrate that our integrated approach outperforms the state-of-the-art. To further improve performance, we will look into strategies that allow us to take advantage even more effectively of our ability to train the network as a whole. In particular, we will look into using hard negative mining strategies over the whole image  instead of relying on pre-extracted patches. This has the potential of producing more discriminative filters and, consequently, better descriptors.
Local Convolutional Features with Unsupervised Training for Image Retrieval.In: ICCV. (2015)