HOTC2026-Modal

HOTC2026-Modal adds human-verified visible/modal segmentation masks and mask-tight modal bounding boxes to all 481 HOTC 2026 videos: 406 training/update videos and 75 Public-Val75 videos, totaling 194,584 annotated frames.

Three official false-color frames are shown for illustration. The middle column shows the HOTC 2026 amodal bounding box. The right column overlays the same amodal box with the HOTC2026-Modal visible/modal segmentation mask and mask-tight modal box.

1. Download and extract the official HOTC 2026 data

The full source-image and hyperspectral datasets are not included in this repository. Download them from either official mirror listed by the HOTC 2026 organizers:

• Google Drive dataset mirror
• Quark Drive dataset mirror

The installer expects the extracted training and validation roots to contain the following modality directories. Additional files and directories are allowed.

HOTC2026/
├── training/
│   ├── HSI-NIR/
│   ├── HSI-NIR-FalseColor/
│   ├── HSI-RedNIR/
│   ├── HSI-RedNIR-FalseColor/
│   ├── HSI-VIS/
│   ├── HSI-VIS-FalseColor_25/
│   └── update/
│       ├── HSI-NIR/
│       ├── HSI-NIR-FalseColor/
│       ├── HSI-RedNIR/
│       ├── HSI-RedNIR-FalseColor/
│       ├── HSI-VIS/
│       └── HSI-VIS-FalseColor/
└── validation/
    ├── HSI-NIR/
    ├── HSI-NIR-FalseColor/
    ├── HSI-RedNIR/
    ├── HSI-RedNIR-FalseColor/
    ├── HSI-VIS/
    └── HSI-VIS-FalseColor/

The two competition CSV files are available from the HOTC 2026 Kaggle data page: 2026training.csv and sample_submisson.csv. The spelling of sample_submisson.csv follows the organizer-provided filename.

mkdir -p data/hotc2026/csv
kaggle competitions download \
  -c hyperspectral-object-tracking-challenge-2026 \
  -p data/hotc2026/csv
unzip data/hotc2026/csv/hyperspectral-object-tracking-challenge-2026.zip \
  -d data/hotc2026/csv

The organizers also provide these supporting resources:

• Evaluation code
• 2D image to hyperspectral cube conversion code
• Python code for X2Cube and evaluation

2. Download HOTC2026-Modal

hf download ryo818/HOTC2026-Modal \
  --repo-type dataset \
  --local-dir data/hotc2026-modal

cd data/hotc2026-modal
sha256sum --check SHA256SUMS
cd ../..

3. Build the integrated dataset

Run the installer with the two extracted official-data roots and the two Kaggle CSVs:

python data/hotc2026-modal/install_annotations.py \
  data/hotc2026-modal \
  --training-root /path/to/HOTC2026/training \
  --validation-root /path/to/HOTC2026/validation \
  --training-csv data/hotc2026/csv/2026training.csv \
  --sample-csv data/hotc2026/csv/sample_submisson.csv \
  --output data/hotc2026-modal-v1 \
  --path-base .

If archives were extracted into separate locations, repeat the corresponding root option for each location.

The installer verifies every archive and mask checksum, matches every annotation to the official CSV identifier, checks all sequence-level frame counts, and confirms that each mask has the same spatial dimensions as its false-color frame. It does not copy the official data; the integrated index references the user's local files.

The resulting dataset layout is:

data/hotc2026-modal-v1/
├── manifest.json
├── frames.jsonl
├── training/<sensor>/<sequence>/annotation.json
├── training/<sensor>/<sequence>/masks/*.png
├── update/<sensor>/<sequence>/...
└── validation/<sensor>/<sequence>/...

frames.jsonl contains one record per frame:

{
  "frame_key": "training/nir-basketball3_1241",
  "frame_number": 1241,
  "collection": "training",
  "sensor": "nir",
  "sequence_name": "basketball3",
  "false_color_image": "path/to/0001.jpg",
  "hsi_mosaic": "path/to/0001.png",
  "mask": "data/hotc2026-modal-v1/training/nir/basketball3/masks/001241.png",
  "official_amodal_bbox_xywh": [172, 191, 16, 15],
  "modal_bbox_xywh": [174, 191, 17, 17]
}

Bounding boxes use integer [x, y, width, height] image coordinates. The official amodal box is null for Public-Val75 because its ground truth is not public. The modal box is tight around visible mask pixels and is null when the target is fully occluded. A mask value greater than zero is foreground.

To extract and verify only the annotation layer, omit all four official-data and CSV arguments:

python data/hotc2026-modal/install_annotations.py \
  data/hotc2026-modal \
  --output data/hotc2026-modal-v1

Files on the Hub

The Hub stores one ZIP per video under data/train_update/ or data/public_val75/. Each ZIP contains only annotation.json and binary PNG masks. Except for the single illustrative README figure above, original images are not packaged. Hyperspectral files, organizer-provided boxes, CSV labels, and leaderboard ground truth are excluded.

Public-Val75 annotations are an author-created annotation set, not organizer leaderboard ground truth. They were excluded from model fitting, model selection, and component evaluation in the accompanying work.

License and citation

The annotation layer and documentation are released under CC BY 4.0. This license does not apply to the official HOTC data; obtain it from the organizer-listed sources and follow the applicable rules and terms.

@inproceedings{yuzawa2026hyperdam,
  title     = {HyperDAM: Hyperspectral Distractor-Aware Memory with Amodal Expansion
               for SAM 3 Tracking},
  author    = {Yuzawa, Ryoga and Takagi, Tasuku},
  booktitle = {2026 IEEE Workshop on Hyperspectral Image and Signal Processing:
               Evolution in Remote Sensing (WHISPERS)},
  year      = {2026}
}