Predicting the visual search target of users is important for a range of applications, particularly human-computer interaction, given that it reveals information about users’ search intents. During visual search, gaze conveys rich information about the target a user has in mind. Recent development of eye tracking technology makes usage of gaze data in different task more affordable and is expected to become more ubiquitous in the future. Using gaze data for recognizing the search target or intent can be considered very convenient for users as gaze is captured implicitly and users don’t need to verbalized or give any implicit input to the machine for assistance.
More broadly, we want to understand to what level of detail such search targets and intends can be recovered. Recently, several studies shown the possibility of visual search target prediction from gazeSattar et al. [2016, 2015], Borji et al. , Zelinsky et al. . Pioneer works tried to predict the search target in the limited set of known targets from gaze dataZelinsky et al. , Borji et al. . More recent works, move smoothly from closed set of targets to an open-world setting Sattar et al. . Most recently, Sattar et al. 
introduced the gaze pooling layer that can recognize categories and attributes of visual search target. They also removed the need of gaze to train a classifier to recognize the search target of users. This enables the usage of gaze in conjuncture with state of the art deep learning recognition system to recognize users intent. Yet, previous efforts on analyzing search targets and intents have only addressed a classification scenario – hence a finite set of discrete classes.
On the other hand, cognitive neuro scientists have shown first success of visualizing images based on fMRI data Cowen et al. , Nishimoto et al. . While we also want to access aspects of the mental state, our task is fundamentally different in two aspects: (1) While they aim at decoding a specific image that was shown to a person, our goal is to decode visual search target. (2) While they are using fMRI data, we are are using gaze data. On a related note, we believe that gaze data is particular interesting to investigate, as it is practical and affordable to collect and use in many application scenarios of future interfaces Sato et al. , Zhang et al. .
As we are targeting decodings of categorical search targets, generating visualization is challenging due to strong intra-class variations. However, recent advances in deep learning have led to a new generation of generative models for images. Recently, Yan et al.  generates images of objects from high-level descriptions. They transfer the high-level text informations to a set of attributes (e.g hair color: Brown, gender: female). These attributes are used later on to build an attribute-conditioned generative model.
Hence, we approach decoding of search targets from gaze data by bringing together recent success in gaze encoding and categorization with state of the art category conditioned generative image models.
The main contributions of this work are:
First proof of concept that visual representations of search targets can be decoded from human gaze data.
We present a practical approach, as it respects the difficulties in collecting large human gaze datasets. Encoder and decoders are trained from large image corpora and transfer between the two representation is facilitated by a semantic layer in between.
We show the importance of localized gaze information for improved search target reconstruction.
2 Related Work
Our work is informed by previous works on gaze-based search target prediction, as well as generative models for image synthesis.
Predicting Search Targets From Gaze
Several previous works focused on predicting the search targets of users from gaze behavior. Zelinsky et al. predicted search targets from subjects’ gaze patterns during a categorical search task Zelinsky et al. . In their experiments, participants were asked to find two categorical search targets among four visually similar distractors. Borji et al. focused on predicting search targets from fixations Borji et al. . They used a binary pattern and 3-level luminance patterns as target and participants were asked to find targets out of a set of other patterns. A compatibility measure was introduced by Sattar et al.  to predict the search targets of users in an open and closed-world setting. In a follow-up work Sattar et al. , they introduced the gaze-pooling layer to integrate gaze data into CNN to predict the category and attributes of search targets. In this work, we are inspired by their work as we also use a gaze-pooling layer as gaze encoder. Yet, our work is first to address reconstruction of the search target from fixation data, which is more challenging than prediction as it addresses a continuous output space.
Image generation and multi-modal learning
Due to recent advances of representation learning and convolution neural networks, image generation becomes possible. Recently, generative adversarial networks (GANs) Goodfellow et al. , Reed et al. , Denton et al. , Isola et al. , Pathak et al.  were used to generate realistic and novel images. GANs consists of two parts: a generator and discriminator. The discriminator is designed to discriminate between generated images and training data. However, training GANS is a challenging task due to the min-max objective. A stochastic variational inference and learning algorithm was introduced by Kingma and Welling 
. A lower bound estimator is achieved via re-parameterization of the variational lower bound. Consequently, standard stochastic gradient method can be used to optimize the estimator. However, the posterior distribution of latent variables is usually unknown. Yang et al. introduced a general-optimization based approach that uses image generation models and latent priors for posterior inferenceYan et al. . They generated images conditioned on visual attributes. In our work, we employ their idea of conditional generative models in the context of inferring search intends from gaze data.
Visual Experience Reconstruction using fMRI
Recent developments in functional magnetic resonance imaging (fMRI) make it possible for neuroscientist to generate links between brain activity and the visual world.
In a more advanced setting, Nishimoto et al. reconstructed natural movies from brain activity Nishimoto et al. . They proposed a motion energy encoding methods to decode the fast visual information and BOLD signals in occipitotemporal visual cortex and fit the model separately to individual voxels.
In another work, Cowen et al. proposed to reconstruct human faces from evoked brain activity using multi-variant regression and PCA Cowen et al. . In their experiment they asked participants to look at an image and then tried to reconstruct this specific image from fMRI data. All of the aboves task tried to reconstruct the seen images. In constrast, our approach decodes the visual search target of user’s which only resides in user’s mind. Also, we are not using fMRI but gaze data which is arguably more practical and affordable to collect and use.
In this work we address decoding of users visual search target from gaze. In order to achieve our goal, we build on recent success in gaze encoding and categorization as well as state of the art category conditioned generative image models. The gaze encoding is used to encode the raw gaze data into a semantic categorical space. The generative image model is conditioned on the encoded gaze data to decode visual search target of users as illustrated in Figure 3.
More formally, participant were searching for a target category in the collage image . During the search task users performed fixations , where each fixation is a triple of positions in screen coordinates and fixation duration . our goal is to sample the visual search target of the target category from
where corresponds to the encoding of the gaze data into a semantic space of and a decoding from that semantic space to the visual search target. In the following, we explain the gaze encoder and search target decoder.
3.1 Semantic Gaze Encoder
following afterSattar et al. , we represent fixations with a Fixation Density Map (FDM):
where each fixation
is represented by a Gaussian spatial distribution of fixed variance at the location of the fixated pointFDM(f). The FDM are then combined with visual features F(I), obtain from a GAP Deep Learning Architecture Zhou et al.  in a Gaze Pooling Layer. The integration is done via element-wise multiplication of FDM and F(I):
To get a final class prediction the weighted feature maps are averaged and fed into a fully connected and softmax layer:
are learned weights and is the bias. We use those class posterior as gaze encoding and condition the generation model on them. We hypothesis that this encoding convey rich information on user’s visual search target.
3.2 Visual Search Target Decoder
Category-Conditioned Image Generation model
Conditional Variational Auto-Encoder (CVAE) Yan et al. 
has been shown successfully sample images based on attribute classes. Here, we build on this recent success and train the model for generation of images of different clothing categories. Given the category vectorand the latent variable , our goal is to build a generative model , which generates image . The generated image is conditioned on the categorical information and the latent variable. In conditional variational auto-encoder, the auxiliary distribution is introduced to approximate the true posterior . The goal of learning process is to find the best parameter which maximizes the lower bound of the log-likelihood Hence the conditional log-likelihood is
where the variational lower bound
is maximized for learning the model parameter.
3.3 Pruning Strategies
We describe the details of our method and the design choices that make our method more robust to noise. After all, human gaze data is noisy and needs aggregation over multiple observations in order to extract information. In the following, we describe how we aggregate information across multiple stimuli as well as prune the representation of the gaze encoder.
Gaze Encoding over Multiple Stimuli
Typical dataset for human gaze contains, data over multiple stimuli. For the dataset that we are going to use in the next section, such stimuli correspond to multiple collages that are shown to the participants. We compute weighted average of posteriors of fixated images with fixation durations to get a prediction over a collage. Moreover, we also average the resulted posteriors per collage to get one final posterior over the set of collages for each categories.
Coping with Noise in Gaze Encoder Prediction
Despite the aggregation as described above, inferred quantities from gaze are typically still noisy. This can also be seen from the prior study using the gaze encoder Sattar et al. . In order to cope with this challenge, we try different pruning strategies to suppress weak activations in the semantic representation .
Specifically, we tried four scenarios to decode visual search target from gaze. In the first case, we used the plain posterior as conditioned vector. In the remaining cases, we used only the top 1 to top 3 highest activated classes in the posterior as condition vector for CVAE. All other probabilities are set to zero and the posterior is renormalized afterwards.
We evaluate our approach on an existing dataset of human gaze. Besides qualitative results, we show two user studies. One measures the success in reconstructing meaningful visual representations of visual search targets and the second highlight the importance of a gaze encoder that respects localized information. We start by describing the dataset we use throughout our experimental evaluation.
We evaluate our model on the gaze dataset from Sattar et al. . Their data set contains gaze data of 14 participants searching for 10 different clothing categories and attributes. They used Deep-fashionZhou et al.  dataset to train their proposed network and synthesizes their collages. Consequently, we also used Deep-fashion data set to train the CVAE for image generation. In addition, we train a classifier (VGGNet-16-GAP) as explained in Sattar et al.  to get class posterior from gaze data. Both networks are trained over top 10 categories of Deep-fashion. We use the same train, test and validation split as proposed in the deep-fashion dataset.
4.2 Qualitative Results of Search Target Decoding
Figure 2, Figure 3, Figure 4 show qualitative results of our approach. As explained in subsection 3.3, noise is inherent in the gaze encoding and we therefore consider two pruning strategies to suppress minor activations. Using directly the posterior of the CNN with gaze pooling layer, causes images that contains several categories rather than the intended visual search target. As shown in Figure 2, the image reconstructed from unaltered posteriors (first row) is more blurry and does not seem to contain one specific category (e.g. Z1 looks like a Blouse, Z6 is a dress and Z9 is a skirt). Images from top2 and top1 appear to be more focused on the intended category. Using top2 posteriors generates images which are a mixture between the two posteriors. One can see more details in top2 which is a composition between dress and skirt, whereas the top1 only contains skirt. Images from top1 are sharper and mostly contain one category. However, if the predicted category is wrong, we can not decode the intended category(last row of Figure 3). Also Sattar et al.  showed strong recognition performance for top 1 and top 2 classification performance. Hence, using the top2 and top3, is likely to contain information about the target. This is reflected in better reconstructions for the top2 and top3 strategies.
As one can see in Figure 3, top1 decoded a dress, although the intended category was tank. The intended category is recovered in the top2 and top3 decodings.
As top2 results are giving images with preferred search target, for further analysis, we chooses decoded images from top2. Figure 4 shows 10 samples for each category using top2 posteriors based on fixations from one user. Our approach is able to generate images of different categories. The model performs better for several classes, as Jean, Shorts, Skirt, Dress, Tank, Sweater and Tee. Images from the cardigan and jacket are more similar to each other, although there are still differences in the appearance. In particular, images depicting the search target cardigan appear more elongated compared to those for jackets.
4.3 User Study: Search Target Recognition
In order to assess the accuracy of the search target reconstruction we run a user study. For each participant and category in the study we show 10 samples from our model. The reconstructions are based on human fixations from the dataset mentioned above. The users are asked to pick one category among 10 categories for a given image. The average accuracy of 19 users across categories was
and the detailed confusion matrix is shown inTable 2. There are confusion between cardigan and jacket, also skirt and blouse with dress. Yet, all search targets were recognized significantly above chance (10%). Users were most confident for jean, shorts and dress.
4.4 User Study: Local Vs Global Gaze Encoding
In this user study, we test the importance of local gaze information for the decoding of visual search targets. Our full model is denoted as “local” here, as it uses the full gaze information – in particular the fixation location on the image. We compare it to a “global” model which uses gaze information – but only to the extend that we know an image was fixated without knowing the exact location. This also connects to the analysis performed on the recognition task for the gaze pooling layer in Sattar et al. . We ask how much of a difference these approaches make in terms of search target reconstruction.
In this user study, each participant saw two rows of search target reconstructions. One row was generate by the local, the other by the global method ( Figure 5. The users were instructed to select the row which matches the best, the given search target category. The users selected the local encoding method in of the cases. The chance level for this experiment is as for each image the participates do a binary task. Consequently, the gain is
. The gain indicates how much performance of users differs from a random selections. Also, we performed Chi-Square Goodness of Fit Test, to investigate the significant of our result. The null hypothesis was that both local and global decoding are equal. Thevalue is 6.914 and P-Value is 0.009. Hence, the result is significant at and therefore local information is key for improved search target reconstruction. Detailed results are shown in Table 1.
In this paper, we introduce the first approach to decode visual search target of users from their gaze data. This task is very challenging as the target only resides in user mind. For this aim we used recent advances in generative image models and search target prediction using gaze data. Proof of concept is established by two user studies, showing that the decoded target lead to human recognizable visual representations as well as highlighting the importance of localized gaze information. We like to emphasize that due to the training setup, the method remains highly practical and applicable, as no large scale gaze data had to be collected or used. Key is rather the utilization of a semantic layer that connects the gaze encoder with the conditional generative image model.
- Sattar et al.  Hosnieh Sattar, Andreas Bulling, and Mario Fritz. Predicting the category and attributes of visual search targets using deep gaze pooling. arXiv, abs/1611.10162, 2016.
- Sattar et al.  Hosnieh Sattar, Sabine Müller, Mario Fritz, and Andreas Bulling. Prediction of search targets from fixations in open-world settings. In CVPR, 2015.
- Borji et al.  Ali Borji, Andreas Lennartz, and Marc Pomplun. What do eyes reveal about the mind?: Algorithmic inference of search targets from fixations. Neurocomputing, 2014.
- Zelinsky et al.  Gregory J Zelinsky, Yifan Peng, and Dimitris Samaras. Eye can read your mind: Decoding gaze fixations to reveal categorical search targets. Journal of Vision, 13(14):10, 2013.
- Cowen et al.  Alan S. Cowen, Marvin M. Chun, and Brice A. Kuhl. Neural portraits of perception: Reconstructing face images from evoked brain activity. NeuroImage, 94:12 – 22, 2014.
- Nishimoto et al.  Shinji Nishimoto, An T. Vu, Thomas Naselaris, Yuval Benjamini, Bin Yu, and Jack L. Gallant. Reconstructing visual experiences from brain activity evoked by natural movies. Current Biology, 21(19):1641 – 1646, 2011.
- Sato et al.  Yoichi Sato, Yusuke Sugano, Akihiro Sugimoto, Yoshinori Kuno, and Hideki Koike. Sensing and controlling human gaze in daily living space for human-harmonized information environments. In Human-Harmonized Information Technology, Volume 1, pages 199–237. Springer Japan, 2016.
- Zhang et al.  Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Appearance-based gaze estimation in the wild. In CVPR, 2015.
- Yan et al.  Xinchen Yan, Jimei Yang, Kihyuk Sohn, and Honglak Lee. Attribute2image: Conditional image generation from visual attributes. In ECCV, 2016.
- Goodfellow et al.  Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, 2014.
- Reed et al.  Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text-to-image synthesis. In ICML, 2016.
- Denton et al.  Emily L Denton, Soumith Chintala, Arthur Szlam, and Rob Fergus. Deep generative image models using a laplacian pyramid of adversarial networks. In NIPS, 2015.
- Isola et al.  Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. arxiv, 2016.
- Pathak et al.  Deepak Pathak, Philipp Krähenbühl, Jeff Donahue, Trevor Darrell, and Alexei Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016.
- Kingma and Welling  Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2013.
Zhou et al. 
Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba.
Learning deep features for discriminative localization.In CVPR, 2016.