mscheidl/htd3

HardTracksDataset: A Benchmark for Robust Object Tracking under Heavy Occlusion and Challenging Conditions

Computer Vision Lab, ETH Zurich

Introduction

We introduce the HardTracksDataset (HTD), a novel multi-object tracking (MOT) benchmark specifically designed to address two critical limitations prevalent in existing tracking datasets. First, most current MOT benchmarks narrowly focus on restricted scenarios, such as pedestrian movements, dance sequences, or autonomous driving environments, thus lacking the object diversity and scenario complexity representative of real-world conditions. Second, datasets featuring broader vocabularies, such as, OVT-B and TAO, typically do not sufficiently emphasize challenging scenarios involving long-term occlusions, abrupt appearance changes, and significant position variations. As a consequence, the majority of tracking instances evaluated are relatively easy, obscuring trackers’ limitations on truly challenging cases. HTD addresses these gaps by curating a challenging subset of scenarios from existing datasets, explicitly combining large vocabulary diversity with severe visual challenges. By emphasizing difficult tracking scenarios, particularly long-term occlusions and substantial appearance shifts, HTD provides a focused benchmark aimed at fostering the development of more robust and reliable tracking algorithms for complex real-world situations.

Results of state of the art trackers on HTD

Method	Validation				Test
	TETA	LocA	AssocA	ClsA	TETA	LocA	AssocA	ClsA
Motion-based
ByteTrack	34.877	54.624	19.085	30.922	37.875	56.135	19.464	38.025
DeepSORT	33.782	57.350	15.009	28.987	37.099	58.766	15.729	36.803
OCSORT	33.012	57.599	12.558	28.880	35.164	59.117	11.549	34.825
Appearance-based
MASA	42.246	60.260	34.241	32.237	43.656	60.125	31.454	39.390
OV-Track	29.179	47.393	25.758	14.385	33.586	51.310	26.507	22.941
Transformer-based
OVTR	26.585	44.031	23.724	14.138	29.771	46.338	24.974	21.643
MASA+	42.716	60.364	35.252	32.532	44.063	60.319	32.735	39.135

Download Instructions

To download the dataset you can use the HuggingFace CLI. First install the HuggingFace CLI according to the official HuggingFace documentation and login with your HuggingFace account. Then, you can download the dataset using the following command:

huggingface-cli download mscheidl/htd --repo-type dataset

The dataset is organized in the following structure:

├── htd
    ├── data
        ├── AnimalTrack
        ├── BDD
        ├── ...
    ├── annotations
        ├── classes.txt
        ├── hard_tracks_dataset_coco_test.json
        ├── hard_tracks_dataset_coco_val.json
        ├── ...
    ├── metadata
        ├── lvis_v1_clip_a+cname.npy
        ├── lvis_v1_train_cat_info.json

The data folder contains the videos, the annotations folder contains the annotations in COCO (TAO) format, and the metadata folder contains the metadata files for running MASA+. If you use HTD independently, you can ignore the metadata folder.

Annotation format for HTD dataset

The annotations folder is structured as follows:

├── annotations
    ├── classes.txt
    ├── hard_tracks_dataset_coco_test.json
    ├── hard_tracks_dataset_coco_val.json
    ├── hard_tracks_dataset_coco.json
    ├── hard_tracks_dataset_coco_class_agnostic.json

Details about the annotations:

• classes.txt: Contains the list of classes in the dataset. Useful for Open-Vocabulary tracking.
• hard_tracks_dataset_coco_test.json: Contains the annotations for the test set.
• hard_tracks_dataset_coco_val.json: Contains the annotations for the validation set.
• hard_tracks_dataset_coco.json: Contains the annotations for the entire dataset.
• hard_tracks_dataset_coco_class_agnostic.json: Contains the annotations for the entire dataset in a class-agnostic format. This means that there is only one category namely "object" and all the objects in the dataset are assigned to this category.

The HTD dataset is annotated in COCO format. The annotations are stored in JSON files, which contain information about the images, annotations, categories, and other metadata. The format of the annotations is as follows:

{
    "images": [image],
    "videos": [video],
    "tracks": [track],
    "annotations": [annotation],
    "categories": [category]
}

image: {
    "id": int,                            # Unique ID of the image
    "video_id": int,                      # Reference to the parent video
    "file_name": str,                     # Path to the image file
    "width": int,                         # Image width in pixels
    "height": int,                        # Image height in pixels
    "frame_index": int,                   # Index of the frame within the video (starting from 0)
    "frame_id": int                       # Redundant or external frame ID (optional alignment)
    "video": str,                         # Name of the video 
    "neg_category_ids": [int],            # List of category IDs explicitly not present (optional)
    "not_exhaustive_category_ids": [int]  # Categories not exhaustively labeled in this image (optional)
        
video: {
    "id": int,                            # Unique video ID
    "name": str,                          # Human-readable or path-based name
    "width": int,                         # Frame width
    "height": int,                        # Frame height
    "neg_category_ids": [int],            # List of category IDs explicitly not present (optional)
    "not_exhaustive_category_ids": [int]  # Categories not exhaustively labeled in this video (optional)
    "frame_range": int,                   # Number of frames between annotated frames
    "metadata": dict,                     # Metadata for the video    
}
        
track: {
    "id": int,             # Unique track ID
    "category_id": int,    # Object category
    "video_id": int        # Associated video
}
        
category: {
    "id": int,            # Unique category ID
    "name": str,          # Human-readable name of the category
}
        
annotation: {
    "id": int,                    # Unique annotation ID
    "image_id": int,              # Image/frame ID
    "video_id": int,              # Video ID
    "track_id": int,              # Associated track ID
    "bbox": [x, y, w, h],         # Bounding box in absolute pixel coordinates
    "area": float,                # Area of the bounding box
    "category_id": int            # Category of the object
    "iscrowd": int,               # Crowd flag (from COCO)
    "segmentation": [],           # Polygon-based segmentation (if available)
    "instance_id": int,           # Instance index with a video
    "scale_category": str         # Scale type (e.g., 'moving-object')
}