Official Training and Inference Code of Amodal Expander, Proposed in Tracking Any Object Amodally.
📙 Project Page | 
Official Github | 📎 Paper Link | ✏️ Citations
Amodal Expander serves as a plug-in module that can “amodalize” any ex-isting detector or tracker with limited (amodal) training data.
📌 Leave a ⭐ in our official repository to keep track of the updates.
See installation instructions.
We augment the SOTA modal tracker GTR with Amodal Expander by fine-tuning on TAO-Amodal dataset.
Please prepare datasets and check our MODEL ZOO for training/inference instructions.
After obtaining the prediction JSON lvis_instances_results.json through the above inference pipeline. You can evaluate the tracker results using our evaluation toolkit.
You can test our model on a single video through:
python demo.py --config-file configs/GTR_TAO_Amodal_Expander_PasteNOcclude.yaml \
--video-input demo/input_video.mp4 \
--output demo/output.mp4 \
--opts MODEL.WEIGHTS /path/to/Amodal_Expander_PnO_45k.pthUse
--input video_folder/*.jpginstead if the video consists of image frames.
PasteNOcclude serves as a data augmentation technique to automatically generate more occlusion scenarios. Check the Jupyter demo and implementation details (link 1, link 2, link 3).
This repository is built upon Global Tracking Transformer and Detectron2.
Check here for further details.
@misc{hsieh2023tracking,
title={Tracking Any Object Amodally},
author={Cheng-Yen Hsieh and Tarasha Khurana and Achal Dave and Deva Ramanan},
year={2023},
eprint={2312.12433},
archivePrefix={arXiv},
primaryClass={cs.CV}
}