European Conference on Computer Vision (ECCV), 2024

We introduce an end-to-end trainable compression framework for 3D Gaussian Splatting (3D-GS). We learn quantized latent representations and parameterized decoders reducing per-point memory for each Gaussian. We introduce a coarse-to-fine progressive training strategy for accelerating training while improving the optimization. We also develop a pruning stage resulting in fewer Gaussians and faster rendering speeds for 3D-GS.

International Conference on Computer Vision (ICCV), 2023

We introduce an end-to-end trainable compression framework for implicit feature grids by maintaining discrete latent representations and parameterized decoders. The training is made differentiable with various tricks such as gumbel reparameterization. We provide extensive experiments on compression benchmarks for a variety of domains such as images, videos, and 3D scenes showing the generalizability of our approach. outperforming a variety of INR works. Our approach scales well to higher resolution data and converges faster than traditional INR approaches.

International Conference on Learning Representations (ICLR), 2023

We propose an end-to-end trainable framework for simultaneously optimizing for model compression and sparsity of deep networks. We construct a joint training objective penalizing the entropy of latent representations of model weights. We introduce priors to increase the structured sparsity in the parameter space outperforming existing state-of-the-art model compression methods while also achieving inference speedups through structured sparsity.

IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

We propose an implicit neural representation (INR) algorithm for videos by fitting separate networks to groups of frames while performing patch-wise prediction. This design shares computation in the spatial and temporal dimension leading to reduced encoding times. The video representation is modeled autoregressively with networks fit on a current group conditioned on the previous group’s network weights. We achieve 12X speedups compared to prior video INR approaches while maintaining encoding quality and compression rate. Additionally, we adapt to video content and naturally scale to longer videos or larger frame resolution.

Under review

We conduct a large scale study to analyze the role of architectures in SSL with over 100 variants of ResNets/MobileNets. We show that there is no one network that performs consistently well across the scenarios. Based on this, we propose to apply NAS for SSL and show that the searched architectures outperform popular handcrafted architectures (ResNet18 and MobileNetV2) while performing competitively with the larger and computationally heavy ResNet50 on major image classification benchmarks.

IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021

We perform an empirical study to investigate the Lottery Ticket Hypothesis (LTH) for deep network pruning in the context of object detection, instance segmentation, and keypoint estimation. Our studies reveal that lottery tickets obtained from Imagenet pretraining do not transfer well to the downstream tasks. We provide guidance on how to find lottery tickets with up to 80% overall sparsity on different sub-tasks without incurring any drop in the performance.

IEEE International Conference on Computer Vision (ICCV), 2021

We propose an approach for discovering and attributing images to previously unseen GAN sources by utilizing only a small labeled set of GAN generated images. Our pipeline consists of multiple stages of network training, Out-Of-Distribution (OOD) detection, and merging and refining clusters. Our approach attributes images to seen GAN sources while also discovering new sources.

IEEE International Conference on Computer Vision (ICCV), 2019 (Oral)

We propose a generalized blur model for dual-lens (DL) setups on smartphones and formulate an algorithm for blind motion deblurring (BMD) for unconstrained camera configurations. The approach employs prior enforcing scene consistent disparities which handles the ill-posedness present in DL-BMD achieving state-of-the-art results for DL setups.

IEEE Transactions on Computational Imaging

We propose an approach for light field (LF) reconstruction from coded images. Our algorithm consists of three stages; recovering the all-in-focus image from the coded image, estimating the disparity maps for all LF views and rendering the LF by warping the all-in-focus image with the disparity maps. We show results on three multiplexing schemes and also achieve high quality LF reconstructions with conventional cameras and no hardware modifications.