Soumava Paul

Sebastian Cavada, Soumava Paul, Tuan-Hung Vu et al.

Previous work has evaluated physics reasoning in foundation models using synthetic or semi-synthetic scenes and visual question-answering tasks. However, these benchmarks emphasize high-level events and lack the visual fidelity required to assess true low-level Newtonian understanding. We introduce NewtPhys, a 4D physically annotated dataset built from multiview images of real-world scenes with physics-grounded simulations. The dataset provides dense, fine-grained annotations across timesteps -- including 3D forces and amodal per-pixel quantities covering physics, tracking, semantics and geometry -- bridging the gap between simplistic synthetic setups and realistic visual complexity. Using NewtPhys, we systematically evaluate 56 VLMs, including 54 open-weight models and 2 closed-source frontier models, and 10 VFMs and reveal limitations in low-level physics reasoning. Beyond benchmarking, our dataset enables future research in physics-grounded vision and the development of next-generation physics-aware evaluations. Code and datasets are available at https://astra-vision.github.io/NewtPhys.

Soumava Paul, Titir Dutta, Aheli Saha et al.

Image retrieval under generalized test scenarios has gained significant momentum in literature, and the recently proposed protocol of Universal Cross-domain Retrieval is a pioneer in this direction. A common practice in any such generalized classification or retrieval algorithm is to exploit samples from many domains during training to learn a domain-invariant representation of data. Such criterion is often restrictive, and thus in this work, for the first time, we explore the generalized retrieval problem in a data-efficient manner. Specifically, we aim to generalize any pre-trained cross-domain retrieval network towards any unknown query domain/category, by means of adapting the model on the test data leveraging self-supervised learning techniques. Toward that goal, we explored different self-supervised loss functions~(for example, RotNet, JigSaw, Barlow Twins, etc.) and analyze their effectiveness for the same. Extensive experiments demonstrate the proposed approach is simple, easy to implement, and effective in handling data-efficient UCDR.

Soumava Paul, Prakhar Kaushik, Ankit Vaidya et al.

We address semantic 3D part segmentation: decomposing objects into parts with meaningful names. While datasets exist with part annotations, their definitions are inconsistent across datasets, limiting robust training. Previous methods produce unlabeled decompositions or retrieve single parts without complete shape annotations. We propose ALIGN-Parts, which formulates part naming as a direct set alignment task. Our method decomposes shapes into partlets - implicit 3D part representations - matched to part descriptions via bipartite assignment. We combine geometric cues from 3D part fields, appearance cues from multi-view vision features, and semantic knowledge from language-model-generated affordance descriptions. Text-alignment loss ensures partlets share embedding space with text, enabling a theoretically open-vocabulary matching setup, given sufficient data. Our efficient and novel, one-shot, 3D part segmentation and naming method finds applications in several downstream tasks, including serving as a scalable annotation engine. As our model supports zero-shot matching to arbitrary descriptions and confidence-calibrated predictions for known categories, with human verification, we create a unified ontology that aligns PartNet, 3DCoMPaT++, and Find3D, consisting of 1,794 unique 3D parts. We introduce two novel metrics appropriate for the named 3D part segmentation task. We also show examples from our newly created TexParts dataset.

Soumava Paul, Prakhar Kaushik, Alan Yuille

Multiview 3D evaluation assumes that the images being scored are observations of one static 3D scene. This assumption can fail in NVS and sparse-view reconstruction: inputs or generated outputs may contain artifacts, outlier frames, repeated views, or noise, yet still receive high 3D consistency scores. Existing reference-based metrics require ground truth, while ground-truth-free metrics such as MEt3R depend on learned reconstruction backbones whose failure modes are poorly characterized. We study this reliability problem by comparing neural reconstruction priors with classical geometric verification. We introduce \benchmark, a controlled robustness benchmark for multiview 3D consistency, and a parametric family that decomposes neural metrics into backbone, residual, and aggregation components. This family recovers MEt3R and yields variants up to $3\times$ more robust. Our analysis shows that VGGT, MASt3R, DUSt3R, and Fast3R can hallucinate dense geometry and cross-view support for unrelated scenes, repeated images, and random noise. We introduce COLMAP-based metrics that use matches, registration, dense support, and reconstruction failure as failure-aware consistency signals. On real NVS outputs and a structured human study, these metrics achieve up to $4\times$ higher correlation with human judgments than MEt3R.

Soumava Paul, Prakhar Kaushik, Alan Yuille

In this work, we introduce a generative approach for pose-free (without camera parameters) reconstruction of 360 scenes from a sparse set of 2D images. Pose-free scene reconstruction from incomplete, pose-free observations is usually regularized with depth estimation or 3D foundational priors. While recent advances have enabled sparse-view reconstruction of large complex scenes (with high degree of foreground and background detail) with known camera poses using view-conditioned generative priors, these methods cannot be directly adapted for the pose-free setting when ground-truth poses are not available during evaluation. To address this, we propose an image-to-image generative model designed to inpaint missing details and remove artifacts in novel view renders and depth maps of a 3D scene. We introduce context and geometry conditioning using Feature-wise Linear Modulation (FiLM) modulation layers as a lightweight alternative to cross-attention and also propose a novel confidence measure for 3D Gaussian splat representations to allow for better detection of these artifacts. By progressively integrating these novel views in a Gaussian-SLAM-inspired process, we achieve a multi-view-consistent 3D representation. Evaluations on the MipNeRF360 and DL3DV-10K benchmark dataset demonstrate that our method surpasses existing pose-free techniques and performs competitively with state-of-the-art posed (precomputed camera parameters are given) reconstruction methods in complex 360 scenes. Our project page provides additional results, videos, and code.

Soumava Paul, Titir Dutta, Soma Biswas

In this work, for the first time, we address the problem of universal cross-domain retrieval, where the test data can belong to classes or domains which are unseen during training. Due to dynamically increasing number of categories and practical constraint of training on every possible domain, which requires large amounts of data, generalizing to both unseen classes and domains is important. Towards that goal, we propose SnMpNet (Semantic Neighbourhood and Mixture Prediction Network), which incorporates two novel losses to account for the unseen classes and domains encountered during testing. Specifically, we introduce a novel Semantic Neighborhood loss to bridge the knowledge gap between seen and unseen classes and ensure that the latent space embedding of the unseen classes is semantically meaningful with respect to its neighboring classes. We also introduce a mix-up based supervision at image-level as well as semantic-level of the data for training with the Mixture Prediction loss, which helps in efficient retrieval when the query belongs to an unseen domain. These losses are incorporated on the SE-ResNet50 backbone to obtain SnMpNet. Extensive experiments on two large-scale datasets, Sketchy Extended and DomainNet, and thorough comparisons with state-of-the-art justify the effectiveness of the proposed model.

Soumava Paul, Gurunath Reddy M, K Sreenivasa Rao et al.

Singing Voice Detection (SVD) has been an active area of research in music information retrieval (MIR). Currently, two deep neural network-based methods, one based on CNN and the other on RNN, exist in literature that learn optimized features for the voice detection (VD) task and achieve state-of-the-art performance on common datasets. Both these models have a huge number of parameters (1.4M for CNN and 65.7K for RNN) and hence not suitable for deployment on devices like smartphones or embedded sensors with limited capacity in terms of memory and computation power. The most popular method to address this issue is known as knowledge distillation in deep learning literature (in addition to model compression) where a large pre-trained network known as the teacher is used to train a smaller student network. Given the wide applications of SVD in music information retrieval, to the best of our knowledge, model compression for practical deployment has not yet been explored. In this paper, efforts have been made to investigate this issue using both conventional as well as ensemble knowledge distillation techniques.

Saneem Ahmed Chemmengath, Soumava Paul, Samarth Bharadwaj et al.

Zero-shot learning (ZSL) algorithms typically work by exploiting attribute correlations to be able to make predictions in unseen classes. However, these correlations do not remain intact at test time in most practical settings and the resulting change in these correlations lead to adverse effects on zero-shot learning performance. In this paper, we present a new paradigm for ZSL that: (i) utilizes the class-attribute mapping of unseen classes to estimate the change in target distribution (target shift), and (ii) propose a novel technique called grouped Adversarial Learning (gAL) to reduce negative effects of this shift. Our approach is widely applicable for several existing ZSL algorithms, including those with implicit attribute predictions. We apply the proposed technique ($g$AL) on three popular ZSL algorithms: ALE, SJE, and DEVISE, and show performance improvements on 4 popular ZSL datasets: AwA2, aPY, CUB and SUN. We obtain SOTA results on SUN and aPY datasets and achieve comparable results on AwA2.