Anurag Mittal

Prashant Kumar, Dhruv Makwana, Onkar Susladkar et al.

Accurate static structure reconstruction and segmentation of non-stationary objects is of vital importance for autonomous navigation applications. These applications assume a LiDAR scan to consist of only static structures. In the real world however, LiDAR scans consist of non-stationary dynamic structures - moving and movable objects. Current solutions use segmentation information to isolate and remove moving structures from LiDAR scan. This strategy fails in several important use-cases where segmentation information is not available. In such scenarios, moving objects and objects with high uncertainty in their motion i.e. movable objects, may escape detection. This violates the above assumption. We present MOVES, a novel GAN based adversarial model that segments out moving as well as movable objects in the absence of segmentation information. We achieve this by accurately transforming a dynamic LiDAR scan to its corresponding static scan. This is obtained by replacing dynamic objects and corresponding occlusions with static structures which were occluded by dynamic objects. We leverage corresponding static-dynamic LiDAR pairs.

Yeruru Asrar Ahmed, Anurag Mittal

Text-to-Image (T2I) synthesis is a challenging task that requires modeling complex interactions between two modalities ( i.e., text and image). A common framework adopted in recent state-of-the-art approaches to achieving such multimodal interactions is to bootstrap the learning process with pre-trained image-aligned text embeddings trained using contrastive loss. Furthermore, these embeddings are typically trained generically and reused across various synthesis models. In contrast, we explore an approach to learning text embeddings specifically tailored to the T2I synthesis network, trained in an end-to-end fashion. Further, we combine generative and contrastive training and use two embeddings, one optimized to enhance the photo-realism of the generated images, and the other seeking to capture text-to-image alignment. A comprehensive set of experiments on three text-to-image benchmark datasets (Oxford-102, Caltech-UCSD, and MS-COCO) reveal that having two separate embeddings gives better results than using a shared one and that such an approach performs favourably in comparison with methods that use text representations from a pre-trained text encoder trained using a discriminative approach. Finally, we demonstrate that such learned embeddings can be used in other contexts as well, such as text-to-image manipulation.

Saikat Dutta, Arulkumar Subramaniam, Anurag Mittal

Video frame interpolation aims to synthesize one or multiple frames between two consecutive frames in a video. It has a wide range of applications including slow-motion video generation, frame-rate up-scaling and developing video codecs. Some older works tackled this problem by assuming per-pixel linear motion between video frames. However, objects often follow a non-linear motion pattern in the real domain and some recent methods attempt to model per-pixel motion by non-linear models (e.g., quadratic). A quadratic model can also be inaccurate, especially in the case of motion discontinuities over time (i.e. sudden jerks) and occlusions, where some of the flow information may be invalid or inaccurate. In our paper, we propose to approximate the per-pixel motion using a space-time convolution network that is able to adaptively select the motion model to be used. Specifically, we are able to softly switch between a linear and a quadratic model. Towards this end, we use an end-to-end 3D CNN encoder-decoder architecture over bidirectional optical flows and occlusion maps to estimate the non-linear motion model of each pixel. Further, a motion refinement module is employed to refine the non-linear motion and the interpolated frames are estimated by a simple warping of the neighboring frames with the estimated per-pixel motion. Through a set of comprehensive experiments, we validate the effectiveness of our model and show that our method outperforms state-of-the-art algorithms on four datasets (Vimeo, DAVIS, HD and GoPro).

Arulkumar Subramaniam, Jayesh Vaidya, Muhammed Abdul Majeed Ameen et al.

Video-based computer vision tasks can benefit from estimation of the salient regions and interactions between those regions. Traditionally, this has been done by identifying the object regions in the images by utilizing pre-trained models to perform object detection, object segmentation and/or object pose estimation. Although using pre-trained models is a viable approach, it has several limitations in the need for an exhaustive annotation of object categories, a possible domain gap between datasets, and a bias that is typically present in pre-trained models. In this work, we propose to utilize the common rationale that a sequence of video frames capture a set of common objects and interactions between them, thus a notion of co-segmentation between the video frame features may equip the model with the ability to automatically focus on task-specific salient regions and improve the underlying task's performance in an end-to-end manner. In this regard, we propose a generic module called ``Co-Segmentation inspired Attention Module'' (COSAM) that can be plugged in to any CNN model to promote the notion of co-segmentation based attention among a sequence of video frame features. We show the application of COSAM in three video-based tasks namely: 1) Video-based person re-ID, 2) Video captioning, & 3) Video action classification and demonstrate that COSAM is able to capture the task-specific salient regions in video frames, thus leading to notable performance improvements along with interpretable attention maps for a variety of video-based vision tasks, with possible application to other video-based vision tasks as well.

Gouthaman KV, Anurag Mittal

Generalizing beyond the experiences has a significant role in developing practical AI systems. It has been shown that current Visual Question Answering (VQA) models are over-dependent on the language-priors (spurious correlations between question-types and their most frequent answers) from the train set and pose poor performance on Out-of-Distribution (OOD) test sets. This conduct limits their generalizability and restricts them from being utilized in real-world situations. This paper shows that the sequence model architecture used in the question-encoder has a significant role in the generalizability of VQA models. To demonstrate this, we performed a detailed analysis of various existing RNN-based and Transformer-based question-encoders, and along, we proposed a novel Graph attention network (GAT)-based question-encoder. Our study found that a better choice of sequence model in the question-encoder improves the generalizability of VQA models even without using any additional relatively complex bias-mitigation approaches.

Sinzith Tatikonda, Athira Nambiar, Anurag Mittal

Face is one of the predominant means of person recognition. In the process of ageing, human face is prone to many factors such as time, attributes, weather and other subject specific variations. The impact of these factors were not well studied in the literature of face aging. In this paper, we propose a novel holistic model in this regard viz., ``Face Age progression With Attribute Manipulation (FAWAM)", i.e. generating face images at different ages while simultaneously varying attributes and other subject specific characteristics. We address the task in a bottom-up manner, as two submodules i.e. face age progression and face attribute manipulation. For face aging, we use an attribute-conscious face aging model with a pyramidal generative adversarial network that can model age-specific facial changes while maintaining intrinsic subject specific characteristics. For facial attribute manipulation, the age processed facial image is manipulated with desired attributes while preserving other details unchanged, leveraging an attribute generative adversarial network architecture. We conduct extensive analysis in standard large scale datasets and our model achieves significant performance both quantitatively and qualitatively.

Saikat Dutta, Nisarg A. Shah, Anurag Mittal

This paper explores an efficient solution for Space-time Super-Resolution, aiming to generate High-resolution Slow-motion videos from Low Resolution and Low Frame rate videos. A simplistic solution is the sequential running of Video Super Resolution and Video Frame interpolation models. However, this type of solutions are memory inefficient, have high inference time, and could not make the proper use of space-time relation property. To this extent, we first interpolate in LR space using quadratic modeling. Input LR frames are super-resolved using a state-of-the-art Video Super-Resolution method. Flowmaps and blending mask which are used to synthesize LR interpolated frame is reused in HR space using bilinear upsampling. This leads to a coarse estimate of HR intermediate frame which often contains artifacts along motion boundaries. We use a refinement network to improve the quality of HR intermediate frame via residual learning. Our model is lightweight and performs better than current state-of-the-art models in REDS STSR Validation set.

Divya Kothandaraman, Athira Nambiar, Anurag Mittal

Practical autonomous driving systems face two crucial challenges: memory constraints and domain gap issues. In this paper, we present a novel approach to learn domain adaptive knowledge in models with limited memory, thus bestowing the model with the ability to deal with these issues in a comprehensive manner. We term this as "Domain Adaptive Knowledge Distillation" and address the same in the context of unsupervised domain-adaptive semantic segmentation by proposing a multi-level distillation strategy to effectively distil knowledge at different levels. Further, we introduce a novel cross entropy loss that leverages pseudo labels from the teacher. These pseudo teacher labels play a multifaceted role towards: (i) knowledge distillation from the teacher network to the student network & (ii) serving as a proxy for the ground truth for target domain images, where the problem is completely unsupervised. We introduce four paradigms for distilling domain adaptive knowledge and carry out extensive experiments and ablation studies on real-to-real as well as synthetic-to-real scenarios. Our experiments demonstrate the profound success of our proposed method.

Rahul Chakwate, Arulkumar Subramaniam, Anurag Mittal

Representation learning from 3D point clouds is challenging due to their inherent nature of permutation invariance and irregular distribution in space. Existing deep learning methods follow a hierarchical feature extraction paradigm in which high-level abstract features are derived from low-level features. However, they fail to exploit different granularity of information due to the limited interaction between these features. To this end, we propose Multi-Abstraction Refinement Network (MARNet) that ensures an effective exchange of information between multi-level features to gain local and global contextual cues while effectively preserving them till the final layer. We empirically show the effectiveness of MARNet in terms of state-of-the-art results on two challenging tasks: Shape classification and Coarse-to-fine grained semantic segmentation. MARNet significantly improves the classification performance by 2% over the baseline and outperforms the state-of-the-art methods on semantic segmentation task.

Vikram Singh, Anurag Mittal

Divide and conquer is an established algorithm design paradigm that has proven itself to solve a variety of problems efficiently. However, it is yet to be fully explored in solving problems with a neural network, particularly the problem of image super-resolution. In this work, we propose an approach to divide the problem of image super-resolution into multiple sub-problems and then solve/conquer them with the help of a neural network. Unlike a typical deep neural network, we design an alternate network architecture that is much wider (along with being deeper) than existing networks and is specially designed to implement the divide-and-conquer design paradigm with a neural network. Additionally, a technique to calibrate the intensities of feature map pixels is being introduced. Extensive experimentation on five datasets reveals that our approach towards the problem and the proposed architecture generate better and sharper results than current state-of-the-art methods.

Gouthaman KV, Athira Nambiar, Kancheti Sai Srinivas et al.

Attention models are widely used in Vision-language (V-L) tasks to perform the visual-textual correlation. Humans perform such a correlation with a strong linguistic understanding of the visual world. However, even the best performing attention model in V-L tasks lacks such a high-level linguistic understanding, thus creating a semantic gap between the modalities. In this paper, we propose an attention mechanism - Linguistically-aware Attention (LAT) - that leverages object attributes obtained from generic object detectors along with pre-trained language models to reduce this semantic gap. LAT represents visual and textual modalities in a common linguistically-rich space, thus providing linguistic awareness to the attention process. We apply and demonstrate the effectiveness of LAT in three V-L tasks: Counting-VQA, VQA, and Image captioning. In Counting-VQA, we propose a novel counting-specific VQA model to predict an intuitive count and achieve state-of-the-art results on five datasets. In VQA and Captioning, we show the generic nature and effectiveness of LAT by adapting it into various baselines and consistently improving their performance.

Gouthaman KV, Anurag Mittal

Recent studies have shown that current VQA models are heavily biased on the language priors in the train set to answer the question, irrespective of the image. E.g., overwhelmingly answer "what sport is" as "tennis" or "what color banana" as "yellow." This behavior restricts them from real-world application scenarios. In this work, we propose a novel model-agnostic question encoder, Visually-Grounded Question Encoder (VGQE), for VQA that reduces this effect. VGQE utilizes both visual and language modalities equally while encoding the question. Hence the question representation itself gets sufficient visual-grounding, and thus reduces the dependency of the model on the language priors. We demonstrate the effect of VGQE on three recent VQA models and achieve state-of-the-art results on the bias-sensitive split of the VQAv2 dataset; VQA-CPv2. Further, unlike the existing bias-reduction techniques, on the standard VQAv2 benchmark, our approach does not drop the accuracy; instead, it improves the performance.

Anubha Pandey, Ashish Mishra, Vinay Kumar Verma et al.

Conventional approaches to Sketch-Based Image Retrieval (SBIR) assume that the data of all the classes are available during training. The assumption may not always be practical since the data of a few classes may be unavailable, or the classes may not appear at the time of training. Zero-Shot Sketch-Based Image Retrieval (ZS-SBIR) relaxes this constraint and allows the algorithm to handle previously unseen classes during the test. This paper proposes a generative approach based on the Stacked Adversarial Network (SAN) and the advantage of Siamese Network (SN) for ZS-SBIR. While SAN generates a high-quality sample, SN learns a better distance metric compared to that of the nearest neighbor search. The capability of the generative model to synthesize image features based on the sketch reduces the SBIR problem to that of an image-to-image retrieval problem. We evaluate the efficacy of our proposed approach on TU-Berlin, and Sketchy database in both standard ZSL and generalized ZSL setting. The proposed method yields a significant improvement in standard ZSL as well as in a more challenging generalized ZSL setting (GZSL) for SBIR.

Manoj Kumar Lenka, Anubha Pandey, Anurag Mittal

Deblurring is the task of restoring a blurred image to a sharp one, retrieving the information lost due to the blur. In blind deblurring we have no information regarding the blur kernel. As deblurring can be considered as an image to image translation task, deep learning based solutions, including the ones which use GAN (Generative Adversarial Network), have been proven effective for deblurring. Most of them have an encoder-decoder structure. Our objective is to try different GAN structures and improve its performance through various modifications to the existing structure for supervised deblurring. In supervised deblurring we have pairs of blurred and their corresponding sharp images, while in the unsupervised case we have a set of blurred and sharp images but their is no correspondence between them. Modifications to the structures is done to improve the global perception of the model. As blur is non-uniform in nature, for deblurring we require global information of the entire image, whereas convolution used in CNN is able to provide only local perception. Deep models can be used to improve global perception but due to large number of parameters it becomes difficult for it to converge and inference time increases, to solve this we propose the use of attention module (non-local block) which was previously used in language translation and other image to image translation tasks in deblurring. Use of residual connection also improves the performance of deblurring as features from the lower layers are added to the upper layers of the model. It has been found that classical losses like L1, L2, and perceptual loss also help in training of GANs when added together with adversarial loss. We also concatenate edge information of the image to observe its effects on deblurring. We also use feedback modules to retain long term dependencies

Sasi Kiran Yelamarthi, Shiva Krishna Reddy, Ashish Mishra et al.

Sketch-based image retrieval (SBIR) is the task of retrieving images from a natural image database that correspond to a given hand-drawn sketch. Ideally, an SBIR model should learn to associate components in the sketch (say, feet, tail, etc.) with the corresponding components in the image having similar shape characteristics. However, current evaluation methods simply focus only on coarse-grained evaluation where the focus is on retrieving images which belong to the same class as the sketch but not necessarily having the same shape characteristics as in the sketch. As a result, existing methods simply learn to associate sketches with classes seen during training and hence fail to generalize to unseen classes. In this paper, we propose a new benchmark for zero-shot SBIR where the model is evaluated in novel classes that are not seen during training. We show through extensive experiments that existing models for SBIR that are trained in a discriminative setting learn only class specific mappings and fail to generalize to the proposed zero-shot setting. To circumvent this, we propose a generative approach for the SBIR task by proposing deep conditional generative models that take the sketch as an input and fill the missing information stochastically. Experiments on this new benchmark created from the "Sketchy" dataset, which is a large-scale database of sketch-photo pairs demonstrate that the performance of these generative models is significantly better than several state-of-the-art approaches in the proposed zero-shot framework of the coarse-grained SBIR task.

Ashish Mishra, Vinay Kumar Verma, M Shiva Krishna Reddy et al.

We present a generative framework for zero-shot action recognition where some of the possible action classes do not occur in the training data. Our approach is based on modeling each action class using a probability distribution whose parameters are functions of the attribute vector representing that action class. In particular, we assume that the distribution parameters for any action class in the visual space can be expressed as a linear combination of a set of basis vectors where the combination weights are given by the attributes of the action class. These basis vectors can be learned solely using labeled data from the known (i.e., previously seen) action classes, and can then be used to predict the parameters of the probability distributions of unseen action classes. We consider two settings: (1) Inductive setting, where we use only the labeled examples of the seen action classes to predict the unseen action class parameters; and (2) Transductive setting which further leverages unlabeled data from the unseen action classes. Our framework also naturally extends to few-shot action recognition where a few labeled examples from unseen classes are available. Our experiments on benchmark datasets (UCF101, HMDB51 and Olympic) show significant performance improvements as compared to various baselines, in both standard zero-shot (disjoint seen and unseen classes) and generalized zero-shot learning settings.

Ashish Mishra, M Shiva Krishna Reddy, Anurag Mittal et al.

Zero shot learning in Image Classification refers to the setting where images from some novel classes are absent in the training data but other information such as natural language descriptions or attribute vectors of the classes are available. This setting is important in the real world since one may not be able to obtain images of all the possible classes at training. While previous approaches have tried to model the relationship between the class attribute space and the image space via some kind of a transfer function in order to model the image space correspondingly to an unseen class, we take a different approach and try to generate the samples from the given attributes, using a conditional variational autoencoder, and use the generated samples for classification of the unseen classes. By extensive testing on four benchmark datasets, we show that our model outperforms the state of the art, particularly in the more realistic generalized setting, where the training classes can also appear at the test time along with the novel classes.

Santhosh K. Ramakrishnan, Swarna Kamlam Ravindran, Anurag Mittal

Traditional point tracking algorithms such as the KLT use local 2D information aggregation for feature detection and tracking, due to which their performance degrades at the object boundaries that separate multiple objects. Recently, CoMaL Features have been proposed that handle such a case. However, they proposed a simple tracking framework where the points are re-detected in each frame and matched. This is inefficient and may also lose many points that are not re-detected in the next frame. We propose a novel tracking algorithm to accurately and efficiently track CoMaL points. For this, the level line segment associated with the CoMaL points is matched to MSER segments in the next frame using shape-based matching and the matches are further filtered using texture-based matching. Experiments show improvements over a simple re-detect-and-match framework as well as KLT in terms of speed/accuracy on different real-world applications, especially at the object boundaries.

Santhosh K. Ramakrishnan, Ambar Pal, Gaurav Sharma et al.

We study the problem of answering questions about images in the harder setting, where the test questions and corresponding images contain novel objects, which were not queried about in the training data. Such setting is inevitable in real world-owing to the heavy tailed distribution of the visual categories, there would be some objects which would not be annotated in the train set. We show that the performance of two popular existing methods drop significantly (up to 28%) when evaluated on novel objects cf. known objects. We propose methods which use large existing external corpora of (i) unlabeled text, i.e. books, and (ii) images tagged with classes, to achieve novel object based visual question answering. We do systematic empirical studies, for both an oracle case where the novel objects are known textually, as well as a fully automatic case without any explicit knowledge of the novel objects, but with the minimal assumption that the novel objects are semantically related to the existing objects in training. The proposed methods for novel object based visual question answering are modular and can potentially be used with many visual question answering architectures. We show consistent improvements with the two popular architectures and give qualitative analysis of the cases where the model does well and of those where it fails to bring improvements.

Arulkumar Subramaniam, Vismay Patel, Ashish Mishra et al.

We propose a novel approach for First Impressions Recognition in terms of the Big Five personality-traits from short videos. The Big Five personality traits is a model to describe human personality using five broad categories: Extraversion, Agreeableness, Conscientiousness, Neuroticism and Openness. We train two bi-modal end-to-end deep neural network architectures using temporally ordered audio and novel stochastic visual features from few frames, without over-fitting. We empirically show that the trained models perform exceptionally well, even after training from a small sub-portions of inputs. Our method is evaluated in ChaLearn LAP 2016 Apparent Personality Analysis (APA) competition using ChaLearn LAP APA2016 dataset and achieved excellent performance.

Sarthak Parui, Anurag Mittal

Proliferation of touch-based devices has made sketch-based image retrieval practical. While many methods exist for sketch-based object detection/image retrieval on small datasets, relatively less work has been done on large (web)-scale image retrieval. In this paper, we present an efficient approach for image retrieval from millions of images based on user-drawn sketches. Unlike existing methods for this problem which are sensitive to even translation or scale variations, our method handles rotation, translation, scale (i.e. a similarity transformation) and small deformations. The object boundaries are represented as chains of connected segments and the database images are pre-processed to obtain such chains that have a high chance of containing the object. This is accomplished using two approaches in this work: a) extracting long chains in contour segment networks and b) extracting boundaries of segmented object proposals. These chains are then represented by similarity-invariant variable length descriptors. Descriptor similarities are computed by a fast Dynamic Programming-based partial matching algorithm. This matching mechanism is used to generate a hierarchical k-medoids based indexing structure for the extracted chains of all database images in an offline process which is used to efficiently retrieve a small set of possible matched images for query chains. Finally, a geometric verification step is employed to test geometric consistency of multiple chain matches to improve results. Qualitative and quantitative results clearly demonstrate superiority of the approach over existing methods.

Smit Marvaniya, Raj Gupta, Anurag Mittal

Matching deformable objects using their shapes is an important problem in computer vision since shape is perhaps the most distinguishable characteristic of an object. The problem is difficult due to many factors such as intra-class variations, local deformations, articulations, viewpoint changes and missed and extraneous contour portions due to errors in shape extraction. While small local deformations has been handled in the literature by allowing some leeway in the matching of individual contour points via methods such as Chamfer distance and Hausdorff distance, handling more severe deformations and articulations has been done by applying local geometric corrections such as similarity or affine. However, determining which portions of the shape should be used for the geometric corrections is very hard, although some methods have been tried. In this paper, we address this problem by an efficient search for the group of contour segments to be clustered together for a geometric correction using Dynamic Programming by essentially searching for the segmentations of two shapes that lead to the best matching between them. At the same time, we allow portions of the contours to remain unmatched to handle missing and extraneous contour portions. Experiments indicate that our method outperforms other algorithms, especially when the shapes to be matched are more complex.

Anoop Katti, Anurag Mittal

Part-based models with restrictive tree-structured interactions for the Human Pose Estimation problem, leaves many part interactions unhandled. Two of the most common and strong manifestations of such unhandled interactions are self-occlusion among the parts and the confusion in the localization of the non-adjacent symmetric parts. By handling the self-occlusion in a data efficient manner, we improve the performance of the basic Mixture of Parts model by a large margin, especially on uncommon poses. Through addressing the confusion in the symmetric limb localization using a combination of two complementing trees, we improve the performance on all the parts by atmost doubling the running time. Finally, we show that the combination of the two solutions improves the results. We report results that are equivalent to the state-of-the-art on two standard datasets. Because of maintaining the tree-structured interactions and only part-level modeling of the base Mixture of Parts model, this is achieved in time that is much less than the best performing part-based model.

Swarna Kamlam Ravindran, Anurag Mittal

Feature or interest points typically use information aggregation in 2D patches which does not remain stable at object boundaries when there is object motion against a significantly varying background. Level or iso-intensity curves are much more stable under such conditions, especially the longer ones. In this paper, we identify stable portions on long iso-curves and detect corners on them. Further, the iso-curve associated with a corner is used to discard portions from the background and improve matching. Such CoMIC (Corners on Maximally-stable Iso-intensity Curves) points yield superior results at the object boundary regions compared to state-of-the-art detectors while performing comparably at the interior regions as well. This is illustrated in exhaustive matching experiments for both boundary and non-boundary regions in applications such as stereo and point tracking for structure from motion in video sequences.