kilinw/vq2d_cvpr

This repo contains the code for the recipe of the winning entry to the Ego4d VQ2D challenge at CVPR 2022.

Methods for Visual Queries 2D Localization

This repo is a codebase for our submission to the VQ2D task in Ego4D Challenge (CVPR22 and ECCV22). The aim of this repo is to help other researchers and challenge practitioners:

• reproduce some of our experiment results and
• leverage our pre-trained detection model for other tasks.

Currently, this codebase supports the following methods:

• Negative Frames Matter in Egocentric Visual Query 2D Localization
• Modeling Object Proposal Sets for Egocentric Visual Query Localization

Updates

• We released our model ckpt and predicted boxes. See instructions in [INSTALL.md].(INSTALL.md)

• Our code to ECCV22 challenge is released!

• Our checkpoints are released! You can also find them in the Ego4D Model Zoo: https://ego4d-data.org/docs/model-zoo/.

  • Config and Checkpoint, trained with VQ2D v1.0 (used in the first challenge)
  • Config and Checkpoint, trained with VQ2D v1.05 (recommended)

Introduction

We deals with the problem of localizing objects in image and video datasets from visual exemplars. In particular, we focus on the challenging problem of egocentric visual query localization. We first identify grave implicit biases in current query-conditioned model design and visual query datasets. Then, we directly tackle such biases at both frame and object set levels. Concretely, our method solves these issues by expanding limited annotations and dynamically dropping object proposals during training. Additionally, we propose a novel transformer-based module that allows for object-proposal set context to be considered while incorporating query information. We name our module Conditioned Contextual Transformer or CocoFormer. Our experiments show the proposed adaptations improve egocentric query detection, leading to a better visual query localization system in both 2D and 3D configurations. Thus, we can improve frame-level detection performance from 26.28% to 31.26% in AP, which correspondingly improves the VQ2D and VQ3D localization scores by significant margins. Our improved context-aware query object detector ranked first and second respectively in the VQ2D and VQ3D tasks in the 2nd Ego4D challenge.

Visulization

[easy] frying pan	[hard] blue bin

Installation

Please find installation instructions in INSTALL.md. It includes system requirement, installation guide, and dataset preperation.

Quick Start

Run evaluate_vq2d_one_query.py with our release checkpoint to quickly see the result.

    python evaluate_vq2d_one_query.py \
        model.config_path=$PWD/checkpoint/train_log/slurm_8gpus_4nodes_cocoformer/output/config.yaml \
        model.checkpoint_path=$PWD/checkpoint/train_log/slurm_8gpus_4nodes_cocoformer/output/model_0064999.pth \
        data.split=val logging.visualize=True logging.save_dir=$PWD/visualizations

Bibtex

Our CVPR22 Challenge report is available on arXiv.

@article{xu2022negative,
  title={Negative Frames Matter in Egocentric Visual Query 2D Localization},
  author={Xu, Mengmeng and Fu, Cheng-Yang and Li, Yanghao and Ghanem, Bernard and Perez-Rua, Juan-Manuel and Xiang, Tao},
  journal={arXiv preprint arXiv:2208.01949},
  year={2022}
}

Our ECCV22 Challenge report is available on arXiv.

@article{xu2022where,
  doi = {10.48550/ARXIV.2211.10528},
  url = {https://arxiv.org/abs/2211.10528},
  author = {Xu, Mengmeng and Li, Yanghao and Fu, Cheng-Yang and Ghanem, Bernard and Xiang, Tao and Perez-Rua, Juan-Manuel},
  title = {Where is my Wallet? Modeling Object Proposal Sets for Egocentric Visual Query Localization},  
  journal={arXiv preprint arXiv:2211.10528},
  year={2022}
}

License

Improved Baseline for Visual Queries 2D Localization is released under the MIT license.

Acknowledgements

This codebase relies on detectron2, Ego4d, and episodic-memory repositories.