Anirudh S Chakravarthy

I'm a Machine Learning Engineer at Cruise, working on multi-task learning and large models for long-tail recognition in autonomous driving. I graduated with my Master of Science in Computer Vision (MSCV) at the Robotics Institute, Carnegie Mellon University (CMU), where I was advised by Prof. Deva Ramanan. My research at CMU focused on extending panoptic segmentation into the open world and discovering novel objects without explicit supervision.

Want to chat? Feel free to reach out!

Email  |  Resume  |  Google Scholar  |  LinkedIn

We present PROFIT, one of the first optimizers that has been specifically designed to operate on converged models that need to be incrementally fine-tuned on a new task/dataset. Unlike standard optimizers like SGD or Adam, which operate with minimal assumptions since the model weights might be randomly initialized, PROFIT takes advantage of the additional structure of a converged model to regularize the optimization process for better results.

We introduce a new dataset and benchmark, YouMVOS, for multi-shot video object segmentation.

We identify mask quality as a bottleneck for video instance segmentation. To overcome this, we propose an attention-based network to propagate missing object instances. Our method significantly outperforms previous state-of-the-art algorithms using the Mask R-CNN backbone, by achieving 36.0% mAP on the YouTube-VIS benchmark.

We propose a network, MRSCAtt (Mars Rover Spatial and Channel Attention), which jointly uses spatial and channel attention to accurately classify images. We use images taken by NASA's Curiosity rover on Mars as a dataset to show the superiority of our approach by achieving state-of-the-art results with 81.53% test set accuracy on the MSL Surface Dataset, outperforming other methods.

Large Models for Long-tail Perception

Multi-task Learning for Long-tail Perception

Video Instance Segmentation

Vital Parameter Estimation

Existing camera pose estimation methods make use of ground-truth odometry as supervision, which may be expensive to obtain. In this work, we train a transformer-based pose estimation network in a self-supervised manner, leveraging advances in monocular depth estimation.

For visual localization, we often have multiple cameras mounted onto a robot which can be used to infer odometry (known as multi-view visual odometry). Existing works either heavily rely on the scene geometry or use complicated networks posing challenges for real-world generalization. In this work, we aim to develop simple yet strong baselines for multi-view visual odometry, by fusing estimates using monocular visual odometry.

Multi-person 3D reconstruction is challenging, yet no prior work aims to disambiguate inter-person occlusions using temporal information. Motivated by this, we leverage optical flow as a cue to improve 3D human pose estimation in crowded scenes.

Generative models such as StyleGAN have shown very promising results. However, while using such GANs for face generation, we often encounter cases of non-photorealistic generations (e.g: artifacts, not face-like, etc.). In this project, we aim to formally establish the existence of such failure modes in GANs.

Source code from Jon Barron