Suhas Lohit

Anoop Cherian, Kuan-Chuan Peng, Suhas Lohit et al.

Recent times have witnessed an increasing number of applications of deep neural networks towards solving tasks that require superior cognitive abilities, e.g., playing Go, generating art, ChatGPT, etc. Such a dramatic progress raises the question: how generalizable are neural networks in solving problems that demand broad skills? To answer this question, we propose SMART: a Simple Multimodal Algorithmic Reasoning Task and the associated SMART-101 dataset, for evaluating the abstraction, deduction, and generalization abilities of neural networks in solving visuo-linguistic puzzles designed specifically for children in the 6--8 age group. Our dataset consists of 101 unique puzzles; each puzzle comprises a picture and a question, and their solution needs a mix of several elementary skills, including arithmetic, algebra, and spatial reasoning, among others. To scale our dataset towards training deep neural networks, we programmatically generate entirely new instances for each puzzle, while retaining their solution algorithm. To benchmark performances on SMART-101, we propose a vision and language meta-learning model using varied state-of-the-art backbones. Our experiments reveal that while powerful deep models offer reasonable performances on puzzles in a supervised setting, they are not better than random accuracy when analyzed for generalization. We also evaluate the recent ChatGPT and other large language models on a subset of SMART-101 and find that while these models show convincing reasoning abilities, the answers are often incorrect.

Sk Miraj Ahmed, Suhas Lohit, Kuan-Chuan Peng et al.

Cost-effective depth and infrared sensors as alternatives to usual RGB sensors are now a reality, and have some advantages over RGB in domains like autonomous navigation and remote sensing. As such, building computer vision and deep learning systems for depth and infrared data are crucial. However, large labeled datasets for these modalities are still lacking. In such cases, transferring knowledge from a neural network trained on a well-labeled large dataset in the source modality (RGB) to a neural network that works on a target modality (depth, infrared, etc.) is of great value. For reasons like memory and privacy, it may not be possible to access the source data, and knowledge transfer needs to work with only the source models. We describe an effective solution, SOCKET: SOurce-free Cross-modal KnowledgE Transfer for this challenging task of transferring knowledge from one source modality to a different target modality without access to task-relevant source data. The framework reduces the modality gap using paired task-irrelevant data, as well as by matching the mean and variance of the target features with the batch-norm statistics that are present in the source models. We show through extensive experiments that our method significantly outperforms existing source-free methods for classification tasks which do not account for the modality gap.

Zachariah Carmichael, Suhas Lohit, Anoop Cherian et al.

Prototypical part neural networks (ProtoPartNNs), namely PROTOPNET and its derivatives, are an intrinsically interpretable approach to machine learning. Their prototype learning scheme enables intuitive explanations of the form, this (prototype) looks like that (testing image patch). But, does this actually look like that? In this work, we delve into why object part localization and associated heat maps in past work are misleading. Rather than localizing to object parts, existing ProtoPartNNs localize to the entire image, contrary to generated explanatory visualizations. We argue that detraction from these underlying issues is due to the alluring nature of visualizations and an over-reliance on intuition. To alleviate these issues, we devise new receptive field-based architectural constraints for meaningful localization and a principled pixel space mapping for ProtoPartNNs. To improve interpretability, we propose additional architectural improvements, including a simplified classification head. We also make additional corrections to PROTOPNET and its derivatives, such as the use of a validation set, rather than a test set, to evaluate generalization during training. Our approach, PIXPNET (Pixel-grounded Prototypical part Network), is the only ProtoPartNN that truly learns and localizes to prototypical object parts. We demonstrate that PIXPNET achieves quantifiably improved interpretability without sacrificing accuracy.

Nithin Gopalakrishnan Nair, Anoop Cherian, Suhas Lohit et al.

Conditional generative models typically demand large annotated training sets to achieve high-quality synthesis. As a result, there has been significant interest in designing models that perform plug-and-play generation, i.e., to use a predefined or pretrained model, which is not explicitly trained on the generative task, to guide the generative process (e.g., using language). However, such guidance is typically useful only towards synthesizing high-level semantics rather than editing fine-grained details as in image-to-image translation tasks. To this end, and capitalizing on the powerful fine-grained generative control offered by the recent diffusion-based generative models, we introduce Steered Diffusion, a generalized framework for photorealistic zero-shot conditional image generation using a diffusion model trained for unconditional generation. The key idea is to steer the image generation of the diffusion model at inference time via designing a loss using a pre-trained inverse model that characterizes the conditional task. This loss modulates the sampling trajectory of the diffusion process. Our framework allows for easy incorporation of multiple conditions during inference. We present experiments using steered diffusion on several tasks including inpainting, colorization, text-guided semantic editing, and image super-resolution. Our results demonstrate clear qualitative and quantitative improvements over state-of-the-art diffusion-based plug-and-play models while adding negligible additional computational cost.

Manish Sharma, Moitreya Chatterjee, Kuan-Chuan Peng et al.

The primary bottleneck towards obtaining good recognition performance in IR images is the lack of sufficient labeled training data, owing to the cost of acquiring such data. Realizing that object detection methods for the RGB modality are quite robust (at least for some commonplace classes, like person, car, etc.), thanks to the giant training sets that exist, in this work we seek to leverage cues from the RGB modality to scale object detectors to the IR modality, while preserving model performance in the RGB modality. At the core of our method, is a novel tensor decomposition method called TensorFact which splits the convolution kernels of a layer of a Convolutional Neural Network (CNN) into low-rank factor matrices, with fewer parameters than the original CNN. We first pretrain these factor matrices on the RGB modality, for which plenty of training data are assumed to exist and then augment only a few trainable parameters for training on the IR modality to avoid over-fitting, while encouraging them to capture complementary cues from those trained only on the RGB modality. We validate our approach empirically by first assessing how well our TensorFact decomposed network performs at the task of detecting objects in RGB images vis-a-vis the original network and then look at how well it adapts to IR images of the FLIR ADAS v1 dataset. For the latter, we train models under scenarios that pose challenges stemming from data paucity. From the experiments, we observe that: (i) TensorFact shows performance gains on RGB images; (ii) further, this pre-trained model, when fine-tuned, outperforms a standard state-of-the-art object detector on the FLIR ADAS v1 dataset by about 4% in terms of mAP 50 score.

Haomeng Zhang, Kuan-Chuan Peng, Suhas Lohit et al.

Open-vocabulary 3D object detection methods are able to localize 3D boxes of classes unseen during training. Despite the name, existing methods rely on user-specified classes both at training and inference. We propose to study Auto-Vocabulary 3D Object Detection (AV3DOD), where the classes are automatically generated for the detected objects without any user input. To this end, we introduce Semantic Score (SS) to evaluate the quality of the generated class names. We then develop a novel framework, AV3DOD, which leverages 2D vision-language models (VLMs) to generate rich semantic candidates through image captioning, pseudo 3D box generation, and feature-space semantics expansion. AV3DOD achieves the state-of-the-art (SOTA) performance on both localization (mAP) and semantic quality (SS) on the ScanNetV2 and SUNRGB-D datasets. Notably, it surpasses the SOTA, CoDA, by 3.48 overall mAP and attains a 24.5% relative improvement in SS on ScanNetV2.

Ibraheem Muhammad Moosa, Suhas Lohit, Ye Wang et al.

Token-level adaptive computation seeks to reduce inference cost by allocating more computation to harder tokens and less to easier ones. However, prior work is primarily evaluated on natural-language benchmarks using task-level metrics, where token-level difficulty is unobservable and confounded with architectural factors, making it unclear whether compute allocation truly aligns with underlying complexity. We address this gap through three contributions. First, we introduce a complexity-controlled evaluation paradigm using algorithmic and synthetic language tasks with parameterized difficulty, enabling direct testing of token-level compute allocation. Second, we propose ANIRA, a unified recurrent Transformer framework that supports per-token variable-depth computation while isolating compute allocation decisions from other model factors. Third, we use this framework to conduct a systematic analysis of token-level adaptive computation across alignment with complexity, generalization, and decision timing. Our results show that compute allocation aligned with task complexity can emerge without explicit difficulty supervision, but such alignment does not imply algorithmic generalization: models fail to extrapolate to unseen input sizes despite allocating additional computation. We further find that early compute decisions rely on static structural cues, whereas online halting more closely tracks algorithmic execution state.

Anoop Cherian, River Doyle, Eyal Ben-Dov et al.

Recent large language models (LLMs) are trained on diverse corpora and tasks, leading them to develop complementary strengths. Multi-agent debate (MAD) has emerged as a popular way to leverage these strengths for robust reasoning, though it has mostly been applied to language-only tasks, leaving its efficacy on multimodal problems underexplored. In this paper, we study MAD for solving vision-and-language reasoning problems. Our setup enables generalizing the debate protocol with heterogeneous experts that possess single- and multi-modal capabilities. To this end, we present Weighted Iterative Society-of-Experts (WISE), a generalized and modular MAD framework that partitions the agents into Solvers, that generate solutions, and Reflectors, that verify correctness, assign weights, and provide natural language feedback. To aggregate the agents' solutions across debate rounds, while accounting for variance in their responses and the feedback weights, we present a modified Dawid-Skene algorithm for post-processing that integrates our two-stage debate model. We evaluate WISE on SMART-840, VisualPuzzles, EvoChart-QA, and a new SMART-840++ dataset with programmatically generated problem instances of controlled difficulty. Our results show that WISE consistently improves accuracy by 2-7% over the state-of-the-art MAD setups and aggregation methods across diverse multimodal tasks and LLM configurations.

Danrui Li, Jiahao Zhang, Bernhard Egger et al.

Assembling objects from parts requires understanding multimodal instructions, linking them to 3D components, and predicting physically plausible 6-DoF motions for each assembly step. Existing datasets focus on simplified scenarios, overlooking shape complexities and assembly trajectories in industrial assemblies. We introduce AssemblyBench, a synthetic dataset of 2,789 industrial objects with multimodal instruction manuals, corresponding 3D part models, and part assembly trajectories. We also propose a transformer-based model, AssemblyDyno, which uses the instructional manual and the 3D shape of each part to jointly predict assembly order and part assembly trajectories. AssemblyDyno outperforms prior works in both assembly pose estimation and trajectory feasibility, where the latter is evaluated by our physics-based simulations.

Haomiao Ni, Bernhard Egger, Suhas Lohit et al.

Text-conditioned image-to-video generation (TI2V) aims to synthesize a realistic video starting from a given image (e.g., a woman's photo) and a text description (e.g., "a woman is drinking water."). Existing TI2V frameworks often require costly training on video-text datasets and specific model designs for text and image conditioning. In this paper, we propose TI2V-Zero, a zero-shot, tuning-free method that empowers a pretrained text-to-video (T2V) diffusion model to be conditioned on a provided image, enabling TI2V generation without any optimization, fine-tuning, or introducing external modules. Our approach leverages a pretrained T2V diffusion foundation model as the generative prior. To guide video generation with the additional image input, we propose a "repeat-and-slide" strategy that modulates the reverse denoising process, allowing the frozen diffusion model to synthesize a video frame-by-frame starting from the provided image. To ensure temporal continuity, we employ a DDPM inversion strategy to initialize Gaussian noise for each newly synthesized frame and a resampling technique to help preserve visual details. We conduct comprehensive experiments on both domain-specific and open-domain datasets, where TI2V-Zero consistently outperforms a recent open-domain TI2V model. Furthermore, we show that TI2V-Zero can seamlessly extend to other tasks such as video infilling and prediction when provided with more images. Its autoregressive design also supports long video generation.

Deepti Hegde, Suhas Lohit, Kuan-Chuan Peng et al.

LiDAR datasets for autonomous driving exhibit biases in properties such as point cloud density, range, and object dimensions. As a result, object detection networks trained and evaluated in different environments often experience performance degradation. Domain adaptation approaches assume access to unannotated samples from the test distribution to address this problem. However, in the real world, the exact conditions of deployment and access to samples representative of the test dataset may be unavailable while training. We argue that the more realistic and challenging formulation is to require robustness in performance to unseen target domains. We propose to address this problem in a two-pronged manner. First, we leverage paired LiDAR-image data present in most autonomous driving datasets to perform multimodal object detection. We suggest that working with multimodal features by leveraging both images and LiDAR point clouds for scene understanding tasks results in object detectors more robust to unseen domain shifts. Second, we train a 3D object detector to learn multimodal object features across different distributions and promote feature invariance across these source domains to improve generalizability to unseen target domains. To this end, we propose CLIX$^\text{3D}$, a multimodal fusion and supervised contrastive learning framework for 3D object detection that performs alignment of object features from same-class samples of different domains while pushing the features from different classes apart. We show that CLIX$^\text{3D}$ yields state-of-the-art domain generalization performance under multiple dataset shifts.

Yuyang Hu, Suhas Lohit, Ulugbek S. Kamilov et al.

Deep learning has achieved some success in addressing the challenge of cloud removal in optical satellite images, by fusing with synthetic aperture radar (SAR) images. Recently, diffusion models have emerged as powerful tools for cloud removal, delivering higher-quality estimation by sampling from cloud-free distributions, compared to earlier methods. However, diffusion models initiate sampling from pure Gaussian noise, which complicates the sampling trajectory and results in suboptimal performance. Also, current methods fall short in effectively fusing SAR and optical data. To address these limitations, we propose Diffusion Bridges for Cloud Removal, DB-CR, which directly bridges between the cloudy and cloud-free image distributions. In addition, we propose a novel multimodal diffusion bridge architecture with a two-branch backbone for multimodal image restoration, incorporating an efficient backbone and dedicated cross-modality fusion blocks to effectively extract and fuse features from synthetic aperture radar (SAR) and optical images. By formulating cloud removal as a diffusion-bridge problem and leveraging this tailored architecture, DB-CR achieves high-fidelity results while being computationally efficient. We evaluated DB-CR on the SEN12MS-CR cloud-removal dataset, demonstrating that it achieves state-of-the-art results.

Sourya Basu, Suhas Lohit, Matthew Brand

Group equivariance is a strong inductive bias useful in a wide range of deep learning tasks. However, constructing efficient equivariant networks for general groups and domains is difficult. Recent work by Finzi et al. (2021) directly solves the equivariance constraint for arbitrary matrix groups to obtain equivariant MLPs (EMLPs). But this method does not scale well and scaling is crucial in deep learning. Here, we introduce Group Representation Networks (G-RepsNets), a lightweight equivariant network for arbitrary matrix groups with features represented using tensor polynomials. The key intuition for our design is that using tensor representations in the hidden layers of a neural network along with simple inexpensive tensor operations can lead to expressive universal equivariant networks. We find G-RepsNet to be competitive to EMLP on several tasks with group symmetries such as O(5), O(1, 3), and O(3) with scalars, vectors, and second-order tensors as data types. On image classification tasks, we find that G-RepsNet using second-order representations is competitive and often even outperforms sophisticated state-of-the-art equivariant models such as GCNNs (Cohen & Welling, 2016a) and E(2)-CNNs (Weiler & Cesa, 2019). To further illustrate the generality of our approach, we show that G-RepsNet is competitive to G-FNO (Helwig et al., 2023) and EGNN (Satorras et al., 2021) on N-body predictions and solving PDEs, respectively, while being efficient.

Deepti Hegde, Suhas Lohit, Kuan-Chuan Peng et al.

Popular representation learning methods encourage feature invariance under transformations applied at the input. However, in 3D perception tasks like object localization and segmentation, outputs are naturally equivariant to some transformations, such as rotation. Using pre-training loss functions that encourage equivariance of features under certain transformations provides a strong self-supervision signal while also retaining information of geometric relationships between transformed feature representations. This can enable improved performance in downstream tasks that are equivariant to such transformations. In this paper, we propose a spatio-temporal equivariant learning framework by considering both spatial and temporal augmentations jointly. Our experiments show that the best performance arises with a pre-training approach that encourages equivariance to translation, scaling, and flip, rotation and scene flow. For spatial augmentations, we find that depending on the transformation, either a contrastive objective or an equivariance-by-classification objective yields best results. To leverage real-world object deformations and motion, we consider sequential LiDAR scene pairs and develop a novel 3D scene flow-based equivariance objective that leads to improved performance overall. We show our pre-training method for 3D object detection which outperforms existing equivariant and invariant approaches in many settings.

Vineet R Shenoy, Suhas Lohit, Hassan Mansour et al.

Camera-based monitoring of vital signs, also known as imaging photoplethysmography (iPPG), has seen applications in driver-monitoring, perfusion assessment in surgical settings, affective computing, and more. iPPG involves sensing the underlying cardiac pulse from video of the skin and estimating vital signs such as the heart rate or a full pulse waveform. Some previous iPPG methods impose model-based sparse priors on the pulse signals and use iterative optimization for pulse wave recovery, while others use end-to-end black-box deep learning methods. In contrast, we introduce methods that combine signal processing and deep learning methods in an inverse problem framework. Our methods estimate the underlying pulse signal and heart rate from facial video by learning deep-network-based denoising operators that leverage deep algorithm unfolding and deep equilibrium models. Experiments show that our methods can denoise an acquired signal from the face and infer the correct underlying pulse rate, achieving state-of-the-art heart rate estimation performance on well-known benchmarks, all with less than one-fifth the number of learnable parameters as the closest competing method.

Xinhao Xiang, Kuan-Chuan Peng, Suhas Lohit et al.

3D object detection plays a crucial role in autonomous systems, yet existing methods are limited by closed-set assumptions and struggle to recognize novel objects and their attributes in real-world scenarios. We propose OVODA, a novel framework enabling both open-vocabulary 3D object and attribute detection with no need to know the novel class anchor size. OVODA uses foundation models to bridge the semantic gap between 3D features and texts while jointly detecting attributes, e.g., spatial relationships, motion states, etc. To facilitate such research direction, we propose OVAD, a new dataset that supplements existing 3D object detection benchmarks with comprehensive attribute annotations. OVODA incorporates several key innovations, including foundation model feature concatenation, prompt tuning strategies, and specialized techniques for attribute detection, including perspective-specified prompts and horizontal flip augmentation. Our results on both the nuScenes and Argoverse 2 datasets show that under the condition of no given anchor sizes of novel classes, OVODA outperforms the state-of-the-art methods in open-vocabulary 3D object detection while successfully recognizing object attributes. Our OVAD dataset is released here: https://doi.org/10.5281/zenodo.16904069 .

Hao Tang, Kevin Ellis, Suhas Lohit et al.

The task of estimating the world model describing the dynamics of a real world process assumes immense importance for anticipating and preparing for future outcomes. For applications such as video surveillance, robotics applications, autonomous driving, etc. this objective entails synthesizing plausible visual futures, given a few frames of a video to set the visual context. Towards this end, we propose ProgGen, which undertakes the task of video frame prediction by representing the dynamics of the video using a set of neuro-symbolic, human-interpretable set of states (one per frame) by leveraging the inductive biases of Large (Vision) Language Models (LLM/VLM). In particular, ProgGen utilizes LLM/VLM to synthesize programs: (i) to estimate the states of the video, given the visual context (i.e. the frames); (ii) to predict the states corresponding to future time steps by estimating the transition dynamics; (iii) to render the predicted states as visual RGB-frames. Empirical evaluations reveal that our proposed method outperforms competing techniques at the task of video frame prediction in two challenging environments: (i) PhyWorld (ii) Cart Pole. Additionally, ProgGen permits counter-factual reasoning and interpretable video generation attesting to its effectiveness and generalizability for video generation tasks.

Naoko Sawada, Pedro Miraldo, Suhas Lohit et al.

Neural implicit surface representation techniques are in high demand for advancing technologies in augmented reality/virtual reality, digital twins, autonomous navigation, and many other fields. With their ability to model object surfaces in a scene as a continuous function, such techniques have made remarkable strides recently, especially over classical 3D surface reconstruction methods, such as those that use voxels or point clouds. However, these methods struggle with scenes that have varied and complex surfaces principally because they model any given scene with a single encoder network that is tasked to capture all of low through high-surface frequency information in the scene simultaneously. In this work, we propose a novel, neural implicit surface representation approach called FreBIS to overcome this challenge. FreBIS works by stratifying the scene based on the frequency of surfaces into multiple frequency levels, with each level (or a group of levels) encoded by a dedicated encoder. Moreover, FreBIS encourages these encoders to capture complementary information by promoting mutual dissimilarity of the encoded features via a novel, redundancy-aware weighting module. Empirical evaluations on the challenging BlendedMVS dataset indicate that replacing the standard encoder in an off-the-shelf neural surface reconstruction method with our frequency-stratified encoders yields significant improvements. These enhancements are evident both in the quality of the reconstructed 3D surfaces and in the fidelity of their renderings from any viewpoint.

Vineet R. Shenoy, Shaoju Wu, Armand Comas et al.

Remote estimation of vital signs enables health monitoring for situations in which contact-based devices are either not available, too intrusive, or too expensive. In this paper, we present a modular, interpretable pipeline for pulse signal estimation from video of the face that achieves state-of-the-art results on publicly available datasets.Our imaging photoplethysmography (iPPG) system consists of three modules: face and landmark detection, time-series extraction, and pulse signal/pulse rate estimation. Unlike many deep learning methods that make use of a single black-box model that maps directly from input video to output signal or heart rate, our modular approach enables each of the three parts of the pipeline to be interpreted individually. The pulse signal estimation module, which we call TURNIP (Time-Series U-Net with Recurrence for Noise-Robust Imaging Photoplethysmography), allows the system to faithfully reconstruct the underlying pulse signal waveform and uses it to measure heart rate and pulse rate variability metrics, even in the presence of motion. When parts of the face are occluded due to extreme head poses, our system explicitly detects such "self-occluded" regions and maintains estimation robustness despite the missing information. Our algorithm provides reliable heart rate estimates without the need for specialized sensors or contact with the skin, outperforming previous iPPG methods on both color (RGB) and near-infrared (NIR) datasets.

Anoop Cherian, Kuan-Chuan Peng, Suhas Lohit et al.

Recent years have seen a significant progress in the general-purpose problem solving abilities of large vision and language models (LVLMs), such as ChatGPT, Gemini, etc.; some of these breakthroughs even seem to enable AI models to outperform human abilities in varied tasks that demand higher-order cognitive skills. Are the current large AI models indeed capable of generalized problem solving as humans do? A systematic analysis of AI capabilities for joint vision and text reasoning, however, is missing in the current scientific literature. In this paper, we make an effort towards filling this gap, by evaluating state-of-the-art LVLMs on their mathematical and algorithmic reasoning abilities using visuo-linguistic problems from children's Olympiads. Specifically, we consider problems from the Mathematical Kangaroo (MK) Olympiad, which is a popular international competition targeted at children from grades 1-12, that tests children's deeper mathematical abilities using puzzles that are appropriately gauged to their age and skills. Using the puzzles from MK, we created a dataset, dubbed SMART-840, consisting of 840 problems from years 2020-2024. With our dataset, we analyze LVLMs power on mathematical reasoning; their responses on our puzzles offer a direct way to compare against that of children. Our results show that modern LVLMs do demonstrate increasingly powerful reasoning skills in solving problems for higher grades, but lack the foundations to correctly answer problems designed for younger children. Further analysis shows that there is no significant correlation between the reasoning capabilities of AI models and that of young children, and their capabilities appear to be based on a different type of reasoning than the cumulative knowledge that underlies children's mathematics and logic skills.

Xinhang Liu, Yu-Wing Tai, Chi-Keung Tang et al.

Extensions of Neural Radiance Fields (NeRFs) to model dynamic scenes have enabled their near photo-realistic, free-viewpoint rendering. Although these methods have shown some potential in creating immersive experiences, two drawbacks limit their ubiquity: (i) a significant reduction in reconstruction quality when the computing budget is limited, and (ii) a lack of semantic understanding of the underlying scenes. To address these issues, we introduce Gear-NeRF, which leverages semantic information from powerful image segmentation models. Our approach presents a principled way for learning a spatio-temporal (4D) semantic embedding, based on which we introduce the concept of gears to allow for stratified modeling of dynamic regions of the scene based on the extent of their motion. Such differentiation allows us to adjust the spatio-temporal sampling resolution for each region in proportion to its motion scale, achieving more photo-realistic dynamic novel view synthesis. At the same time, almost for free, our approach enables free-viewpoint tracking of objects of interest - a functionality not yet achieved by existing NeRF-based methods. Empirical studies validate the effectiveness of our method, where we achieve state-of-the-art rendering and tracking performance on multiple challenging datasets.

David W. Romero, Suhas Lohit

Group Convolutional Neural Networks (G-CNNs) constrain learned features to respect the symmetries in the selected group, and lead to better generalization when these symmetries appear in the data. If this is not the case, however, equivariance leads to overly constrained models and worse performance. Frequently, transformations occurring in data can be better represented by a subset of a group than by a group as a whole, e.g., rotations in $[-90^{\circ}, 90^{\circ}]$. In such cases, a model that respects equivariance $\textit{partially}$ is better suited to represent the data. In addition, relevant transformations may differ for low and high-level features. For instance, full rotation equivariance is useful to describe edge orientations in a face, but partial rotation equivariance is better suited to describe face poses relative to the camera. In other words, the optimal level of equivariance may differ per layer. In this work, we introduce $\textit{Partial G-CNNs}$: G-CNNs able to learn layer-wise levels of partial and full equivariance to discrete, continuous groups and combinations thereof as part of training. Partial G-CNNs retain full equivariance when beneficial, e.g., for rotated MNIST, but adjust it whenever it becomes harmful, e.g., for classification of 6 / 9 digits or natural images. We empirically show that partial G-CNNs pair G-CNNs when full equivariance is advantageous, and outperform them otherwise.

Suhas Lohit, Shubhendu Trivedi

Omnidirectional images and spherical representations of $3D$ shapes cannot be processed with conventional 2D convolutional neural networks (CNNs) as the unwrapping leads to large distortion. Using fast implementations of spherical and $SO(3)$ convolutions, researchers have recently developed deep learning methods better suited for classifying spherical images. These newly proposed convolutional layers naturally extend the notion of convolution to functions on the unit sphere $S^2$ and the group of rotations $SO(3)$ and these layers are equivariant to 3D rotations. In this paper, we consider the problem of unsupervised learning of rotation-invariant representations for spherical images. In particular, we carefully design an autoencoder architecture consisting of $S^2$ and $SO(3)$ convolutional layers. As 3D rotations are often a nuisance factor, the latent space is constrained to be exactly invariant to these input transformations. As the rotation information is discarded in the latent space, we craft a novel rotation-invariant loss function for training the network. Extensive experiments on multiple datasets demonstrate the usefulness of the learned representations on clustering, retrieval and classification applications.

Suhas Lohit, Michael Jones

Model compression methods are important to allow for easier deployment of deep learning models in compute, memory and energy-constrained environments such as mobile phones. Knowledge distillation is a class of model compression algorithm where knowledge from a large teacher network is transferred to a smaller student network thereby improving the student's performance. In this paper, we show how optimal transport-based loss functions can be used for training a student network which encourages learning student network parameters that help bring the distribution of student features closer to that of the teacher features. We present image classification results on CIFAR-100, SVHN and ImageNet and show that the proposed optimal transport loss functions perform comparably to or better than other loss functions.

Huan Wang, Suhas Lohit, Michael Jones et al.

Several methods of knowledge distillation have been developed for neural network compression. While they all use the KL divergence loss to align the soft outputs of the student model more closely with that of the teacher, the various methods differ in how the intermediate features of the student are encouraged to match those of the teacher. In this paper, we propose a simple-to-implement method using auxiliary classifiers at intermediate layers for matching features, which we refer to as multi-head knowledge distillation (MHKD). We add loss terms for training the student that measure the dissimilarity between student and teacher outputs of the auxiliary classifiers. At the same time, the proposed method also provides a natural way to measure differences at the intermediate layers even though the dimensions of the internal teacher and student features may be different. Through several experiments in image classification on multiple datasets we show that the proposed method outperforms prior relevant approaches presented in the literature.

Huan Wang, Suhas Lohit, Mike Jones et al.

Knowledge distillation (KD) is a general neural network training approach that uses a teacher model to guide the student model. Existing works mainly study KD from the network output side (e.g., trying to design a better KD loss function), while few have attempted to understand it from the input side. Especially, its interplay with data augmentation (DA) has not been well understood. In this paper, we ask: Why do some DA schemes (e.g., CutMix) inherently perform much better than others in KD? What makes a "good" DA in KD? Our investigation from a statistical perspective suggests that a good DA scheme should reduce the covariance of the teacher-student cross-entropy. A practical metric, the stddev of teacher's mean probability (T. stddev), is further presented and well justified empirically. Besides the theoretical understanding, we also introduce a new entropy-based data-mixing DA scheme, CutMixPick, to further enhance CutMix. Extensive empirical studies support our claims and demonstrate how we can harvest considerable performance gains simply by using a better DA scheme in knowledge distillation.

Suhas Lohit, Rushil Anirudh, Pavan Turaga

Motion capture (mocap) and time-of-flight based sensing of human actions are becoming increasingly popular modalities to perform robust activity analysis. Applications range from action recognition to quantifying movement quality for health applications. While marker-less motion capture has made great progress, in critical applications such as healthcare, marker-based systems, especially active markers, are still considered gold-standard. However, there are several practical challenges in both modalities such as visibility, tracking errors, and simply the need to keep marker setup convenient wherein movements are recorded with a reduced marker-set. This implies that certain joint locations will not even be marked-up, making downstream analysis of full body movement challenging. To address this gap, we first pose the problem of reconstructing the unmarked joint data as an ill-posed linear inverse problem. We recover missing joints for a given action by projecting it onto the manifold of human actions, this is achieved by optimizing the latent space representation of a deep autoencoder. Experiments on both mocap and Kinect datasets clearly demonstrate that the proposed method performs very well in recovering semantics of the actions and dynamics of missing joints. We will release all the code and models publicly.

Rushil Anirudh, Suhas Lohit, Pavan Turaga

In this paper, we propose the generative patch prior (GPP) that defines a generative prior for compressive image recovery, based on patch-manifold models. Unlike learned, image-level priors that are restricted to the range space of a pre-trained generator, GPP can recover a wide variety of natural images using a pre-trained patch generator. Additionally, GPP retains the benefits of generative priors like high reconstruction quality at extremely low sensing rates, while also being much more generally applicable. We show that GPP outperforms several unsupervised and supervised techniques on three different sensing models -- linear compressive sensing with known, and unknown calibration settings, and the non-linear phase retrieval problem. Finally, we propose an alternating optimization strategy using GPP for joint calibration-and-reconstruction which performs favorably against several baselines on a real world, un-calibrated compressive sensing dataset.

Suhas Lohit, Qiao Wang, Pavan Turaga

Many time-series classification problems involve developing metrics that are invariant to temporal misalignment. In human activity analysis, temporal misalignment arises due to various reasons including differing initial phase, sensor sampling rates, and elastic time-warps due to subject-specific biomechanics. Past work in this area has only looked at reducing intra-class variability by elastic temporal alignment. In this paper, we propose a hybrid model-based and data-driven approach to learn warping functions that not just reduce intra-class variability, but also increase inter-class separation. We call this a temporal transformer network (TTN). TTN is an interpretable differentiable module, which can be easily integrated at the front end of a classification network. The module is capable of reducing intra-class variance by generating input-dependent warping functions which lead to rate-robust representations. At the same time, it increases inter-class variance by learning warping functions that are more discriminative. We show improvements over strong baselines in 3D action recognition on challenging datasets using the proposed framework. The improvements are especially pronounced when training sets are smaller.

Suhas Lohit, Rajhans Singh, Kuldeep Kulkarni et al.

In resource-constrained environments, one can employ spatial multiplexing cameras to acquire a small number of measurements of a scene, and perform effective reconstruction or high-level inference using purely data-driven neural networks. However, once trained, the measurement matrix and the network are valid only for a single measurement rate (MR) chosen at training time. To overcome this drawback, we answer the following question: How can we jointly design the measurement operator and the reconstruction/inference network so that the system can operate over a \textit{range} of MRs? To this end, we present a novel training algorithm, for learning \textbf{\textit{rate-adaptive}} networks. Using standard datasets, we demonstrate that, when tested over a range of MRs, a rate-adaptive network can provide high quality reconstruction over a the entire range, resulting in up to about 15 dB improvement over previous methods, where the network is valid for only one MR. We demonstrate the effectiveness of our approach for sample-efficient object tracking where video frames are acquired at dynamically varying MRs. We also extend this algorithm to learn the measurement operator in conjunction with image recognition networks. Experiments on MNIST and CIFAR-10 confirm the applicability of our algorithm to different tasks.

Li-Chi Huang, Kuldeep Kulkarni, Anik Jha et al.

Visual Question Answering (VQA) is a complex semantic task requiring both natural language processing and visual recognition. In this paper, we explore whether VQA is solvable when images are captured in a sub-Nyquist compressive paradigm. We develop a series of deep-network architectures that exploit available compressive data to increasing degrees of accuracy, and show that VQA is indeed solvable in the compressed domain. Our results show that there is nominal degradation in VQA performance when using compressive measurements, but that accuracy can be recovered when VQA pipelines are used in conjunction with state-of-the-art deep neural networks for CS reconstruction. The results presented yield important implications for resource-constrained VQA applications.

Suhas Lohit, Pavan Turaga

Non-Euclidean constraints are inherent in many kinds of data in computer vision and machine learning, typically as a result of specific invariance requirements that need to be respected during high-level inference. Often, these geometric constraints can be expressed in the language of Riemannian geometry, where conventional vector space machine learning does not apply directly. The central question this paper deals with is: How does one train deep neural nets whose final outputs are elements on a Riemannian manifold? To answer this, we propose a general framework for manifold-aware training of deep neural networks -- we utilize tangent spaces and exponential maps in order to convert the proposed problem into a form that allows us to bring current advances in deep learning to bear upon this problem. We describe two specific applications to demonstrate this approach: prediction of probability distributions for multi-class image classification, and prediction of illumination-invariant subspaces from a single face-image via regression on the Grassmannian. These applications show the generality of the proposed framework, and result in improved performance over baselines that ignore the geometry of the output space. In addition to solving this specific problem, we believe this paper opens new lines of enquiry centered on the implications of Riemannian geometry on deep architectures.

Suhas Lohit, Kuldeep Kulkarni, Ronan Kerviche et al.

Traditional algorithms for compressive sensing recovery are computationally expensive and are ineffective at low measurement rates. In this work, we propose a data driven non-iterative algorithm to overcome the shortcomings of earlier iterative algorithms. Our solution, ReconNet, is a deep neural network, whose parameters are learned end-to-end to map block-wise compressive measurements of the scene to the desired image blocks. Reconstruction of an image becomes a simple forward pass through the network and can be done in real-time. We show empirically that our algorithm yields reconstructions with higher PSNRs compared to iterative algorithms at low measurement rates and in presence of measurement noise. We also propose a variant of ReconNet which uses adversarial loss in order to further improve reconstruction quality. We discuss how adding a fully connected layer to the existing ReconNet architecture allows for jointly learning the measurement matrix and the reconstruction algorithm in a single network. Experiments on real data obtained from a block compressive imager show that our networks are robust to unseen sensor noise. Finally, through an experiment in object tracking, we show that even at very low measurement rates, reconstructions using our algorithm possess rich semantic content that can be used for high level inference.

Kuldeep Kulkarni, Suhas Lohit, Pavan Turaga et al.

The goal of this paper is to present a non-iterative and more importantly an extremely fast algorithm to reconstruct images from compressively sensed (CS) random measurements. To this end, we propose a novel convolutional neural network (CNN) architecture which takes in CS measurements of an image as input and outputs an intermediate reconstruction. We call this network, ReconNet. The intermediate reconstruction is fed into an off-the-shelf denoiser to obtain the final reconstructed image. On a standard dataset of images we show significant improvements in reconstruction results (both in terms of PSNR and time complexity) over state-of-the-art iterative CS reconstruction algorithms at various measurement rates. Further, through qualitative experiments on real data collected using our block single pixel camera (SPC), we show that our network is highly robust to sensor noise and can recover visually better quality images than competitive algorithms at extremely low sensing rates of 0.1 and 0.04. To demonstrate that our algorithm can recover semantically informative images even at a low measurement rate of 0.01, we present a very robust proof of concept real-time visual tracking application.