TAO-Amodal Dataset

Official Source for Downloading the TAO-Amodal Dataset.

📙 Project Page | 💻 Code | 📎 Paper Link | ✏️ Citations

Dataset Description

Our dataset augments the TAO dataset with amodal bounding box annotations for fully invisible, out-of-frame, and occluded objects. Note that this implies TAO-Amodal also includes modal segmentation masks (as visualized in the color overlays above). Our dataset encompasses 880 categories, aimed at assessing the occlusion reasoning capabilities of current trackers through the paradigm of Tracking Any Object with Amodal perception (TAO-Amodal).

Dataset Download

Download all the annotations.

git lfs install
git clone git@hf.co:datasets/chengyenhsieh/TAO-Amodal

Download all the video frames:

You can either download the frames following the instructions here (recommended) or modify our provided script and run

bash download_TAO.sh

📚 Dataset Structure

The dataset should be structured like this:

   TAO-Amodal
    ├── frames
    │    └── train
    │       ├── ArgoVerse
    │       ├── BDD
    │       ├── Charades
    │       ├── HACS
    │       ├── LaSOT
    │       └── YFCC100M
    ├── amodal_annotations
    │    ├── train/validation/test.json
    │    ├── train_lvis_v1.json
    │    └── validation_lvis_v1.json
    ├── example_output
    │    └── prediction.json
    ├── BURST_annotations
    │    ├── train
    │         └── train_visibility.json
    │    ...

📚 File Descriptions

File Name	Description
train/validation/test.json	Formal annotation files. We use these annotations for visualization. Categories include those in lvis v0.5 and freeform categories.
train_lvis_v1.json	We use this file to train our amodal-expander, treating each image frame as an independent sequence. Categories are aligned with those in lvis v1.0.
validation_lvis_v1.json	We use this file to evaluate our amodal-expander. Categories are aligned with those in lvis v1.0.
prediction.json	Example output json from amodal-expander. Tracker predictions should be structured like this file to be evaluated with our evaluation toolkit.
BURST_annotations/XXX.json	Modal mask annotations from BURST dataset with our heuristic visibility attributes. We provide these files for the convenience of visualization

Annotation and Prediction Format

Our annotations are structured similarly as TAO with some modifications. Annotations:


Annotation file format:
{
    "info" : info,
    "images" : [image],
    "videos": [video],
    "tracks": [track],
    "annotations" : [annotation],
    "categories": [category],
    "licenses" : [license],
}
annotation: {
    "id": int,
    "image_id": int,
    "track_id": int,
    "bbox": [x,y,width,height],
    "area": float,

    # Redundant field for compatibility with COCO scripts
    "category_id": int,
    "video_id": int,

    # Other important attributes for evaluation on TAO-Amodal
    "amodal_bbox": [x,y,width,height],
    "amodal_is_uncertain": bool,
    "visibility": float, (0.~1.0)
}
image, info, video, track, category, licenses, : Same as TAO

Predictions should be structured as:

[{
    "image_id" : int,
    "category_id" : int,
    "bbox" : [x,y,width,height],
    "score" : float,
    "track_id": int,
    "video_id": int
}]

Refer to the instructions of TAO dataset for further details

📺 Example Sequences

Check here for more examples and here for visualization code.

Citation

@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}
}

task_categories:

object-detection
multi-object-tracking

Datasets:

chengyenhsieh
/

TAO-Amodal