Abhinav Agarwalla

I am a Machine Learning Research Engineer at Neural Magic.

Previously, I worked at Argo AI as a Research Engineer. Before that, I was a Masters student at the Robotics Institute, Carnegie Mellon University, where I worked with Prof. Deva Ramanan. Before that, I worked at Video Analytics Lab, Indian Institute of Science, on computer vision problems and at Microsoft, improving search at Bing scale. I graduated from Indian Institute of Technology Kharagpur majoring in Mathematics and Computing.

I have enjoyed working on a diverse set of topic - ranging from video summarisation, motion planning for mobile robots to applying machine learning for pressing issues such as estimating insulin intake, detecting tumorous cells, etc.

Email  /  CV  /  Google Scholar  /  LinkedIn  /  Github

I am excited about developing state-of-the-art algorithms and formulating relevant research problems that enablerobots to sense and perceive the world as humans do. Presently, my research is focused on autonomous driving andpoint clouds, unsupervised learning, domain adaptation and transfer learning

Reviewer: CVPR 2022-2024, ICRA 2021, 2023-2024, NeurIPS 2023 DGM4H, NeurIPS 2021 ICBINB, MLRC 2020, 2022.

The paper introduces a detection-centric network for lidar panoptic segmentation (LPS) and tracking on 3D Lidar point clouds, challenging the conventional "bottom-up" approach. The network utilizes trajectory-level point supervision to obtain `modal` annotations. These are then utilized to predict fine-grained instance segments across time.

Existing self-supervised approaches can learn good representations, but require some labeled data to map these features to the end task of crowd density estimation. We mitigate this issue with the proposed paradigm of complete self-supervision, which does not need even a single labeled image.

Existing unsupervised methods require some annotated samples to facilitate the final task-specific predictions. Instead, we leverage the distribution of labels for supervisory signal such that no image-label pair is needed for training a classifier.

We propose a universal framework to handle both task and domain shifts, and report state-of-the-art results on single/multi-source domain adaptation and domain generalisation.

We integrate learning and motion planning for soccer playing differential drive robots using Bayesian optimisation.

In this paper, we introduce Key-Value Memory Networks to a multimodal setting and a novel key-addressing mechanism to deal with sequence-to-sequence models. The proposed model naturally decomposes the problem of video captioning into vision and language segments, dealing with them as key-value pairs.

CNNs have become the preferred choice for most computer vision tasks. However, these are not best suited for multi-gigapixel resolution Whole Slide Images (WSIs) of histology slides due to large size of these images. This work address this issue with novel 2D-LSTM + CNN network for tumor segmentation.

Fork of a fork :)