TAO-Amodal: A Benchmark for Tracking Any Object Amodally

Cheng-Yen (Wesley) Hsieh1, Kaihua Chen1, Achal Dave2, Tarasha Khurana1, Deva Ramanan1
1Carnegie Mellon University   2Toyota Research Institute  
Paper Code TAO-Amodal Dataset Qualitative Results

TAO-Amodal dataset features diverse (880 categories) annotations for both
Traditional tracking (top) and Amodal tracking (bottom).


Abstract

Amodal perception, the ability to comprehend complete object structures from partial visibility, is a fundamental skill, even for infants. Its significance extends to applications like autonomous driving, where a clear understanding of heavily occluded objects is essential. However, modern detection and tracking algorithms often overlook this critical capability, perhaps due to the prevalence of modal annotations in most benchmarks. To address the scarcity of amodal benchmarks, we introduce TAO-Amodal, featuring 833 diverse categories in thousands of video sequences. Our dataset includes amodal and modal bounding boxes for visible and partially or fully occluded objects, including those that are partially out of the camera frame. We investigate the current lay of the land in both amodal tracking and detection by benchmarking state-of-the-art modal trackers and amodal segmentation methods. We find that existing methods, even when adapted for amodal tracking, struggle to detect and track objects under heavy occlusion. To mitigate this, we explore simple finetuning schemes that can increase the amodal tracking and detection metrics of occluded objects by 2.1% and 3.3%.


Traditional vs. Amodal Tracking

Traditional perception (top) concentrates on identifying visible segments. Consequently, they face peculiar outputs such as vanishing bounding boxes or tiny box sizes under occlusion scenarios. Amodal perception (bottom) advances beyond conventional approaches by inferring complete object boundaries, even when certain portions are occluded.

Teaser





TAO-Amodal Dataset

More examples (amodal) 🤗 Download

Our dataset augments the TAO dataset with amodal bounding box annotations for fully invisible, out-of-frame, and occluded objects across 880 categories. Note that this implies TAO-Amodal also includes modal segmentation masks.










Amodal Expander

Our Amodal Expander serves as a plug-in module that can ``amodalize" any existing detector or tracker with limited (amodal) training data. Here we provide qualitative results of both modal (top) and amodal (bottom) predictions from amodal expander.






Acknowledgements

The data collection efforts behind TAO dataset are crucial for the realization of TAO-Amodal. We also thank BURST dataset for its collection of modal mask annotations. Amodal annotations for this dataset were provided by AnnotateX. We thank Neehar Peri and Jason Zhang from CMU for their detailed feedback on the dataset and experiments.

BibTeX

@misc{hsieh2023tracking,
          title={TAO-Amodal: A Benchmark for Tracking Any Object Amodally},
          author={Cheng-Yen Hsieh and Kaihua Chen and Achal Dave and Tarasha Khurana and Deva Ramanan},
          year={2023},
          eprint={2312.12433},
          archivePrefix={arXiv},
          primaryClass={cs.CV}
      }