Skip to content

Latest commit

 

History

History
75 lines (68 loc) · 2.96 KB

README.md

File metadata and controls

75 lines (68 loc) · 2.96 KB

MessyTable: Instance Association in Multiple Camera Views

Useful Links:

  • Visit our [Project Homepage] for an overview of MessyTable dataset
  • Read our Paper (Accepted in ECCV 2020) [Preprint] for complete technical details

Setup

Environment

(ananconda3)
conda create -n mt python=3.7
conda activate mt
pip install torch==1.1.0 torchvision==0.3.0
pip install opencv-python==3.4.2.17
pip install scipy==1.2.0

Note Python 3.7 is needed to use the KMSolver module we provide in src/, which is a python wrapper of a C++ implementation.

Dataset Preparation

  • Download MessyTable.zip (~22 GB) from [Aliyun] or [Google Drive]
  • Unzip MessyTable.zip, check the unzipped folder includes images/ and labels/
  • Rename the unzipped folder to data/, place data/ in this repository as follows:
MessyTable
├── models
├── src
├── data
    ├── images
    ├── labels

Pretrained Model

  • Download pretrained ASNet (ASNet.pth) from [Google Drive]
  • Place model in models/asnet/ for evaluation

Get started

Evaluation

This example evaluates pretrained ASNet:

python test.py --config_dir asnet \
--eval_json test.json \
--save_features \
--eval_model

Arguments:

  • --config_dir: the directory that contains the specific config file train.yaml (checkpoints are automatically saved in the same dir)
  • --eval_json: data split name in data/labels/ to evaluate (test.json, val.json and by difficulty levels, see Paper Sec 5.3)
  • --save_features: (optional) save extrated features in models/<config_dir> for faster evaluation in the future
  • --load_features: (optional) load saved features from models/<config_dir>, if the features have been saved in the past
  • --eval_model: evaluate using the appearance features only
  • --eval_model_esc: evaluate using the appearance features with epipolar soft constraint (See Paper Sec 5.2)
  • --eval_by_angle: (optional) evaluate by angle differences (See Paper Sec 5.3)

Training

  • Training on 8 GPUs with batch size 512 (we use this setting in our paper)
python train.py --config_dir asnet
  • Training on a single GPU with batch size 64
python train.py --config_dir asnet_1gpu

Citation

If you find this repo or dataset useful, please consider citing our paper

@inproceedings{
    CaiZhang2020MessyTable,
    title={MessyTable: Instance Association in Multiple Camera Views},
    author={Zhongang Cai and Junzhe Zhang and Daxuan Ren and Cunjun Yu and Haiyu Zhao and Shuai Yi and Chai Kiat Yeo and Chen Change Loy},
    booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
    month = {August},
    year={2020}
}