I completed my PhD under the supervision of Prof. Peter V. Gehler at MPI and joined as a research scientist at Nvidia Research. My work lies at the intersection of Machine Learning and Computer Vision. Specifically, I am working on leveraging machine learning techniques for better inference in computer vision applications. The main research question is how to make use of learning techniques such as deep neural networks and random forests for inference in structured prediction frameworks.
In German Conference on Pattern Recognition (GCPR), October 2018 (inproceedings)
Most of the top performing action recognition methods use optical flow as a "black box" input. Here we take a deeper look at the combination of flow and action recognition, and investigate why optical flow is helpful, what makes a flow method good for action recognition, and how we can make it better. In particular, we investigate the impact of different flow algorithms and input transformations to better understand how these affect a state-of-the-art action recognition method. Furthermore, we fine tune two neural-network flow methods end-to-end on the most widely used action recognition dataset (UCF101). Based on these experiments, we make the following five observations: 1) optical flow is useful for action recognition because it is invariant to appearance, 2) optical flow methods are optimized to minimize end-point-error (EPE), but the EPE of current methods is not well correlated with action recognition performance, 3) for the flow methods tested, accuracy at boundaries and at small displacements is most correlated with action recognition performance, 4) training optical flow to minimize classification error instead of minimizing EPE improves recognition performance, and 5) optical flow learned for the task of action recognition differs from traditional optical flow especially inside the human body and at the boundary of the body. These observations may encourage optical flow researchers to look beyond EPE as a goal and guide action recognition researchers to seek better motion cues, leading to a tighter integration of the optical flow and action recognition communities.
We address the unsupervised learning of several interconnected problems in low-level vision: single view depth prediction, camera motion estimation, optical flow and segmentation of a video into the static scene and moving regions. Our key insight is that these four fundamental vision problems are coupled and, consequently, learning to solve them together simplifies the problem because the solutions can reinforce each other by exploiting known geometric constraints. In order to model geometric constraints, we introduce Adversarial Collaboration, a framework that facilitates competition and collaboration between neural networks. We go beyond previous work by exploiting geometry more explicitly and segmenting the scene into static and moving regions. Adversarial Collaboration works much like expectation-maximization but with neural networks that act as adversaries, competing to explain pixels that correspond to static or moving regions, and as collaborators through a moderator that assigns pixels to be either static or independently moving. Our novel method integrates all these problems in a common framework and simultaneously reasons about the segmentation of the scene into moving objects and the static background, the camera motion, depth of the static scene structure, and the optical flow of moving objects. Our model is trained without any supervision and achieves state of the art results amongst unsupervised methods.
German Conference on Pattern Recognition (GCPR), 2018 (conference)
Most of the top performing action recognition methods use optical flow as a black box input. Here we take a deeper look at the combination of flow and action recognition, and investigate why optical flow is helpful, what makes a flow method good for action recognition, and how we can make it better. In particular, we investigate the impact of different flow algorithms and input transformations to better understand how these affect a state-of-the-art action recognition method. Furthermore, we fine tune two neural-network flow methods end-to-end on the most widely used action recognition dataset (UCF101). Based on these experiments, we make the following five observations: 1) optical flow is useful for action recognition because it is invariant to appearance, 2) optical flow methods are optimized to minimize end-point-error (EPE), but the EPE of current methods is not well correlated with action recognition performance, 3) for the flow methods tested, accuracy at boundaries and at small displacements is most correlated with action recognition performance, 4) training optical flow to minimize classification error instead of minimizing EPE improves recognition performance, and 5) optical flow learned for the task of action recognition differs from traditional optical flow especially inside the human body and at the boundary of the body. These observations may encourage optical flow researchers to look beyond EPE as a goal and guide action recognition researchers to seek better motion cues, leading to a tighter integration of the optical flow and action recognition communities.
In Proceedings IEEE International Conference on Computer Vision (ICCV), IEEE, Piscataway, NJ, USA, IEEE International Conference on Computer Vision (ICCV), October 2017 (inproceedings) Accepted
In this work, we propose a technique to convert CNN
models for semantic segmentation of static images into
CNNs for video data. We describe a warping method that
can be used to augment existing architectures with very lit-
tle extra computational cost. This module is called Net-
Warp and we demonstrate its use for a range of network
architectures. The main design principle is to use optical
flow of adjacent frames for warping internal network repre-
sentations across time. A key insight of this work is that fast
optical flow methods can be combined with many different
CNN architectures for improved performance and end-to-
end training. Experiments validate that the proposed ap-
proach incurs only little extra computational cost, while im-
proving performance, when video streams are available. We
achieve new state-of-the-art results on the standard CamVid
and Cityscapes benchmark datasets and show reliable im-
provements over different baseline networks. Our code and
models are available at http://segmentation.is.
In Proceedings IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2017, IEEE, Piscataway, NJ, USA, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017 (inproceedings)
MPI for Intelligent Systems and University of Tübingen, 2017 (phdthesis)
Computer vision can be understood as the ability to perform 'inference' on image data. Breakthroughs in computer vision technology are often marked by advances in inference techniques, as even the model design is often dictated by the complexity of inference in them. This thesis proposes learning based inference schemes and demonstrates applications in computer vision. We propose techniques for inference in both generative and discriminative computer vision models.
Despite their intuitive appeal, the use of generative models in vision is hampered by the difficulty of posterior inference, which is often too complex or too slow to be practical. We propose techniques for improving inference in two widely used techniques: Markov Chain Monte Carlo (MCMC) sampling and message-passing inference. Our inference strategy is to learn separate discriminative models that assist Bayesian inference in a generative model. Experiments on a range of generative vision models show that the proposed techniques accelerate the inference process and/or converge to better solutions.
A main complication in the design of discriminative models is the inclusion of prior knowledge in a principled way. For better inference in discriminative models, we propose techniques that modify the original model itself, as inference is simple evaluation of the model. We concentrate on convolutional neural network (CNN) models and propose a generalization of standard spatial convolutions, which are the basic building blocks of CNN architectures, to bilateral convolutions. First, we generalize the existing use of bilateral filters and then propose new neural network architectures with learnable
bilateral filters, which we call `Bilateral Neural Networks'. We show how the bilateral filtering modules can be used for modifying existing CNN architectures for better image segmentation and propose a neural network approach for temporal information propagation in videos. Experiments demonstrate the potential of the proposed bilateral networks on a wide range of vision tasks and datasets.
In summary, we propose learning based techniques for better inference in several computer vision models ranging from inverse graphics to freely parameterized neural networks. In generative vision models, our inference techniques alleviate some of the crucial hurdles in Bayesian posterior inference, paving new ways for the use of model based machine learning in vision. In discriminative CNN models, the proposed filter generalizations aid in the design of new neural network architectures that can handle sparse high-dimensional data as well as provide a way for incorporating prior knowledge into CNNs.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017 (article)
This paper introduces a fast and efficient segmentation technique for 2D images and 3D point clouds of building facades. Facades of buildings are highly structured and consequently most methods that have been proposed for this problem aim to make use of this strong prior information. Contrary to most prior work, we are describing a system that is almost domain independent and consists of standard segmentation methods. We train a sequence of boosted decision trees using auto-context features. This is learned using stacked generalization. We find that this technique performs better, or comparable with all previous published methods and present empirical results on all available 2D and 3D facade benchmark datasets. The proposed method is simple to implement, easy to extend, and very efficient at test-time inference.
In European Conference on Computer Vision (ECCV), Lecture Notes in Computer Science, Springer, 14th European Conference on Computer Vision, October 2016 (inproceedings)
In this paper we propose a CNN architecture for semantic image segmentation. We introduce a new “bilateral inception” module that can be inserted in existing CNN architectures and performs bilateral filtering, at multiple feature-scales, between superpixels in an image. The feature spaces for bilateral filtering and other parameters of the module are learned end-to-end using standard backpropagation techniques. The bilateral inception module addresses two issues that arise with general CNN segmentation architectures. First, this module propagates information between (super) pixels while respecting image edges, thus using the structured information of the problem for improved results. Second, the layer recovers a full resolution segmentation result from the lower resolution solution of a CNN. In the experiments, we modify several existing CNN architectures by inserting our inception modules between the last CNN (1 × 1 convolution) layers. Empirical results on three different datasets show reliable improvements not only in comparison to the baseline networks, but also in comparison to several dense-pixel prediction techniques such as CRFs, while being competitive in time.
In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages: 3889-3898, IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), June 2016 (inproceedings)
Existing optical flow methods make generic, spatially homogeneous, assumptions about the spatial structure of the flow. In reality, optical flow varies across an image depending on object class.
Simply put, different objects move differently. Here we exploit recent advances in static semantic scene segmentation to segment the image into objects of different types. We define different models of image motion in these regions depending on the type of object. For example, we model the motion on roads with homographies, vegetation with spatially smooth flow, and independently moving objects like cars and planes with affine motion plus deviations. We then pose the flow estimation problem using a novel formulation of localized layers, which addresses limitations of traditional layered models for dealing with complex scene motion. Our semantic flow method achieves the lowest error of any published monocular method in the KITTI-2015 flow benchmark and produces qualitatively better flow and segmentation than recent top methods on a wide range of natural videos.
In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages: 4452-4461, IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), June 2016 (inproceedings)
Bilateral filters have wide spread use due to their edge-preserving properties. The common use case is to manually choose a parametric filter type, usually a Gaussian filter. In this paper, we will generalize the parametrization and in particular derive a gradient descent algorithm so the filter parameters can be learned from data. This derivation allows to learn high dimensional linear filters that operate in sparsely populated feature spaces. We build on the permutohedral lattice construction for efficient filtering. The ability to learn more general forms of high-dimensional filters can be used in several diverse applications. First, we demonstrate the use in applications where single filter applications are desired for runtime reasons. Further, we show how this algorithm can be used to learn the pairwise potentials in densely connected conditional random fields and apply these to different image segmentation tasks. Finally, we introduce layers of bilateral filters in CNNs and propose bilateral neural networks for the use of high-dimensional sparse data. This view provides new ways to encode model structure into network architectures. A diverse set of experiments empirically validates the usage of general forms of filters.
In Special Issue on Generative Models in Computer Vision and Medical Imaging, 136, pages: 32-44, Elsevier, July 2015 (inproceedings)
Computer vision is hard because of a large variability in lighting, shape, and texture; in addition the image signal is non-additive due to occlusion. Generative models promised to account for this variability by accurately modelling the image formation process as a function of latent variables with prior beliefs. Bayesian posterior inference could then, in principle, explain the observation. While intuitively appealing, generative models for computer vision have largely failed to deliver on that promise due to the difficulty of posterior inference. As a result the community has favored efficient discriminative approaches. We still believe in the usefulness of generative models in computer vision, but argue that we need to leverage existing discriminative or even heuristic computer vision methods.
We implement this idea in a principled way in our informed sampler and in careful experiments demonstrate it on challenging models which contain renderer programs as their components. The informed sampler, using simple discriminative proposals based on existing computer vision technology achieves dramatic improvements in inference. Our approach enables a new richness in generative models that was out of reach with existing inference technology.
In ICLR Workshop Track, ICLR, May 2015 (inproceedings)
This paper presents a convolutional layer that is able to process sparse input features. As an example, for image recognition problems this allows an efficient filtering of signals that do not lie on a dense grid (like pixel position), but of more general features (such as color values). The presented algorithm makes use of the permutohedral lattice data structure. The permutohedral lattice was introduced to efficiently implement a bilateral filter, a commonly used image processing operation. Its use allows for a generalization of the convolution type found in current (spatial) convolutional network architectures.
Jampani, V., Eslami, S. M. A., Tarlow, D., Kohli, P., Winn, J.
In Eighteenth International Conference on Artificial Intelligence and Statistics (AISTATS), 38, pages: 425-433, JMLR Workshop and Conference Proceedings, Eighteenth International Conference on Artificial Intelligence and Statistics, May 2015 (inproceedings)
Generative models provide a powerful framework for probabilistic reasoning. However, in many domains their use has been hampered by the practical difficulties of inference. This is particularly the case in computer vision, where models of the imaging process tend to be large, loopy and layered. For this reason bottom-up conditional models have traditionally dominated in such domains. We find that widely-used, general-purpose message passing inference algorithms such as Expectation Propagation (EP) and Variational Message Passing (VMP) fail on the simplest of vision models. With these models in mind, we introduce a modification to message passing that learns to exploit their layered structure by passing 'consensus' messages that guide inference towards good solutions. Experiments on a variety of problems show that the proposed technique leads to significantly more accurate inference results, not only when compared to standard EP and VMP, but also when compared to competitive bottom-up conditional models.
psJampani, V., Eslami, S. M. A., Tarlow, D., Kohli, P., Winn, J.
Consensus Message Passing for Layered Graphical Models
In Eighteenth International Conference on Artificial Intelligence and Statistics (AISTATS), 38, pages: 425-433, JMLR Workshop and Conference Proceedings, Eighteenth International Conference on Artificial Intelligence and Statistics, May 2015 (inproceedings)
In Applications of Computer Vision (WACV), 2015 IEEE Winter Conference on, pages: 1038-1045, IEEE, WACV,, January 2015 (inproceedings)
In this paper we propose a system for the problem of facade segmentation. Building facades are highly structured images and consequently most methods that have been proposed for this problem, aim to make use of this strong prior information. We are describing a system that is almost domain independent and consists of standard segmentation methods. A sequence of boosted decision trees is stacked using auto-context features and learned using the stacked generalization technique. We find that this, albeit standard, technique performs better, or equals, all previous published empirical results on all available facade benchmark datasets. The proposed method is simple to implement, easy to extend, and very efficient at test time inference.
IIIT-Hyderabad, Hyderabad, India, January 2013 (mastersthesis)
Pneumoconiosis is an occupational lung disease caused by the inhalation of industrial dust. Despite the increasing safety measures and better work place environments, pneumoconiosis is deemed to be the most common occupational disease in the developing countries like India and China. Screening and assessment of this disease is done through radiological observation of chest x-rays. Several studies have shown the significant inter and intra reader observer variation in the diagnosis of this disease, showing the complexity of the task and importance of the expertise in diagnosis.
The present study is aimed at understanding the perceptual and cognitive factors affecting the reading of chest x-rays of pneumoconiosis patients. Understanding these factors helps in developing better image acquisition systems, better training regimen for radiologists and development of better computer aided diagnostic (CAD) systems. We used an eye tracking experiment to study the various factors affecting the assessment of this diffused lung disease. Specifically, we aimed at understanding the role of expertize, contralateral symmetric (CS) information present in chest x-rays on the diagnosis and the eye movements of the observers. We also studied the inter and intra observer fixation consistency along with the role of anatomical and bottom up saliency features in attracting the gaze of observers of different expertize levels, to get better insights into the effect of bottom up and top down visual saliency on the eye movements of observers.
The experiment is conducted in a room dedicated to eye tracking experiments. Participants consisting of novices (3), medical students (12), residents (4) and staff radiologists (4) were presented with good quality PA chest X-rays, and were asked to give profusion ratings for each of the 6 lung zones. Image set consisting of 17 normal full chest x-rays and 16 single lung images are shown to the participants in random order. Time of the diagnosis and the eye movements are also recorded using a remote head free eye tracker.
Results indicated that Expertise and CS play important roles in the diagnosis of pneumoconiosis. Novices and medical students are slow and inefficient whereas, residents and staff are quick and efficient. A key finding of our study is that the presence of CS information alone does not help improve diagnosis as much as learning how to use the information. This learning appears to be gained from focused training and years of experience. Hence, good training for radiologists and careful observation of each lung zone may improve the quality of diagnostic results. For residents, the eye scanning strategies play an important role in using the CS information present in chest radiographs; however, in staff radiologists, peripheral vision or higher-level cognitive processes seems to play role in using the CS information.
There is a reasonably good inter and intra observer fixation consistency suggesting the use of similar viewing strategies. Experience is helping the observers to develop new visual strategies based on the image content so that they can quickly and efficiently assess the disease level. First few fixations seem to be playing an important role in choosing the visual strategy, appropriate for the given image.
Both inter-rib and rib regions are given equal importance by the observers. Despite reading of chest x-rays being highly task dependent, bottom up saliency is shown to have played an important role in attracting the fixations of the observers. This role of bottom up saliency seems to be more in lower expertize groups compared to that of higher expertize groups. Both bottom up and top down influence of visual fixations seems to change with time. The relative role of top down and bottom up influences of visual attention is still not completely understood and it remains the part of future work.
Based on our experimental results, we have developed an extended saliency model by combining the bottom up saliency and the saliency of lung regions in a chest x-ray. This new saliency model performed significantly better than bottom-up saliency in predicting the gaze of the observers in our experiment. Even though, the model is a simple combination of bottom-up saliency maps and segmented lung masks, this demonstrates that even basic models using simple image features can predict the fixations of the observers to a good accuracy.
Experimental analysis suggested that the factors affecting the reading of chest x-rays of pneumoconiosis are complex and varied. A good understanding of these factors definitely helps in the development of better radiological screening of pneumoconiosis through improved training and also through the use of improved CAD tools. The presented work is an attempt to get insights into what these factors are and how they modify the behavior of the observers.
Proceedings of the Eighth Indian Conference on Computer Vision, Graphics and Image Processing, pages: 80:1-80:8, ACM, Mumbai, India, December 2012 (conference)
Visual attention plays a major role in our lives. Our very perception (which very much decides our survival) depends on it - like perceiving a predator while walking through a forest, perceiving a fast car coming from the front on a busy road or even spotting our favorite color out of the many colors. In Medical Imaging, where medical experts have to take major clinical decisions based on the examination of images of various kinds (CT, MRI etc), visual attention plays a pivotal role. It makes the medical experts fixate on any abnormal behavior exhibited in the medical image and helps in speedy diagnosis. Many previous works (see the paper for details) have exhibited this important fact and the model proposed by Nodine and Kundel highlights the important role of visual attention in medical image diagnosis.
Visual attention involves two components - Bottom-Up and Top-Down.In the present work, we examine a number of established computational models of visual attention in the context of chest X-rays (infected with Pneumoconiosis) and retinal images (having hard exudates). The fundamental motivation is to try to understand the applicability of visual attention models in the context of different types of abnormalities.
Our assessment of four popular visual attention models, is extensive and shows that they are able to pick up abnormal features reasonably well. We compare the models towards detecting subtle abnormalities and high-contrast lesions. Although significant scope of improvements exists especially in picking up more subtle abnormalities and getting more selective towards picking up more abnormalities and less normal structures, the presented assessment shows that visual attention indeed shows a promise for inclusion in the main field of medical image analysis
In Proc. SPIE 7966, Medical Imaging: Image Perception, Observer Performance, and Technology Assessment, 2011, Florida, March 2011 (inproceedings)
Pneumoconiosis, a lung disease caused by the inhalation of dust, is mainly diagnosed using chest radiographs. The effects of using contralateral symmetric (CS) information present in chest radiographs in the diagnosis of pneumoconiosis are studied using an eye tracking experimental study. The role of expertise and the influence of CS information on the performance of readers with different expertise level are also of interest. Experimental subjects ranging from novices & medical students to staff radiologists were presented with 17 double and 16 single lung images, and were asked to give profusion ratings for each lung zone. Eye movements and the time for their diagnosis were also recorded. Kruskal-Wallis test (χ2(6) = 13.38, p = .038), showed that the observer error (average sum of absolute differences) in double lung images differed significantly across the different expertise categories when considering all the participants. Wilcoxon-signed rank test indicated that the observer error was significantly higher for single-lung images (Z = 3.13, p < .001) than for the double-lung images for all the participants. Mann-Whitney test (U = 28, p = .038) showed that the differential error between single and double lung images is significantly higher in doctors [staff & residents] than in non-doctors [others]. Thus, Expertise & CS information plays a significant role in the diagnosis of pneumoconiosis. CS information helps in diagnosing pneumoconiosis by reducing the general tendency of giving less profusion ratings. Training and experience appear to play important roles in learning to use the CS information present in the chest radiographs.
MSR-TR-2010-148, Microsoft Research, Redmond, 2010 (techreport)
Traditional grid and list representations of image search results are the dominant interaction paradigms that users face on a daily basis, yet it is unclear that such paradigms are well-suited for experiences where the user‟s task is to browse images for leisure, to discover new information or to seek particular images to represent ideas. We introduce ImageFlow, a novel image search user interface that ex-plores a different alternative to the traditional presentation of image search results. ImageFlow presents image results on a canvas where we map semantic features (e.g., rele-vance, related queries) to the canvas‟ spatial dimensions (e.g., x, y, z) in a way that allows for several levels of en-gagement – from passively viewing a stream of images, to seamlessly navigating through the semantic space and ac-tively collecting images for sharing and reuse. We have implemented our system as a fully functioning prototype, and we report on promising, preliminary usage results.
Our goal is to understand the principles of Perception, Action and Learning in autonomous systems that successfully interact with complex environments and to use this understanding to design future systems