Bhavya Goyal

Bhavya Goyal

pronunciation

PhD Candidate
Computer Science UW-Madison
Email: bhavya at cs.wisc.edu

I am a PhD Candidate in CS at UW-Madison.

I am part of the Wision Lab where I am advised by Prof. Mohit Gupta. My research revolves around the areas of Computer Vision, Generative AI, and 3D Sensing. Currently, I am working on 3D inference models under adverse conditions like low light, motion, and LiDAR noise.

Earlier, I worked as a Research Engineer at Visual Understanding Lab of Samsung Research, Seoul for 3 years. I also spent a summer at Cruise AI with the Perception team and Prof. YongJae Lee. Before that, I completed my undergraduate in CS from IIT Delhi where I was advised by Prof. Vinay Riberio.

Publications

  • Probabilisitic Point Clouds
    Robust 3D Object Detection using Probabilistic Point Clouds
    Under Review
    Bhavya Goyal,  F. Barragan,  W. Lin,  A. Velten,  Y. Li,  M. Gupta
  • Dual Corruptions
    Robust Scene Inference under Noise-Blur Dual Corruptions
    ICCP 2022
    Bhavya Goyal,  J.F. Lalonde,  Y. Li,  M. Gupta
  • Photon Net
    Photon-Starved Scene Inference using Single Photon Cameras
    ICCV 2021
    Bhavya Goyal,  M. Gupta
  • ABE
    Attention-based Ensemble for Deep Metric Learning
    ECCV 2018
    W. Kim,  Bhavya Goyal,  K. Chawla,  J. Lee,  K. Kwon

Experience


  • Samsung Research, Seoul

    (Sep'16 - Jul'19)

    Research Engineer (Visual Understanding Lab)

    Grocery
    Object Recognition and Retrieval, Smart Refrigerators [Link]
    • Image recognition algorithms for detecting grocery items inside the refrigerator, models used for product recommendation engine in Samsung Smart Refrigerators
    • Designed techniques using global and attentive deep local descriptors for feature matching, paired with geometric verification of selected keypoints, to recognize products that are partially occluded by other items.

    Vision
    Product Search, Bixby Vision [Link]
    • Developed large scale retrieval models for products in online shopping mall images.
    • Designed attention mechanism in Deep Neural Networks to ignore background noise in query images, achieves SOTA results on all image major retrieval benchmarks, published in ECCV 2018


  • Cruise, San Francisco

    (May'22 - Aug'22)

    Research Intern (Perception)

    SSL
    Self-Supervised Learning
    • Self-supervised pretraining for 3D object recognition models using camera RGB images and LiDAR point clouds.
    • Joint pre-text tasks for 2D images and point clouds based on masked auto-encoders using vision transformers.




Projects

  • AIMeetsBeauty

    AI Meets Beauty Challenge [ Link ] [ Code ]

    ACM Multimedia Conference 2018
    • Winner with SOTA results for half million product image recognition
    • Developed CNN based retrieval model using attention module to ignore background clutter in product images, approx nearest neighbor search for product images in large scale retrieval DB
  • Tiger Re-ID

    Tiger ReID in Wild [ Code ] [ Slides ] [ Paper ]
    Prof. Yin Li

    • Proposed architecture using object detection and re-id which encourages diversity among feature embeddings to get more discriminative features which boosts the performance on most retrieval benchmarks.
  • Stack Exchange

    Stack-Exchange Tag Prediction [ Code ] [ Slides ] [ Paper ]
    Prof. Mausam

    • State-of-the-art results for predicting tags/labels for questions on different Stack Exchange portals
    • Classification of meta features from text and code snippets in questions using SVM
    • KMeans to cluster word embeddings from Google's pre-trained Word2Vec model and ensembled with model trained with term affinity of tags and words
-->