--- pretty_name: "MMOT" language: - en license: cc-by-nc-nd-4.0 task_categories: - multi object tracking tags: - computer-vision - video - multi-object-tracking - drone - uav - aerial-tracking - multispectral - multispectral-imagery - oriented-bounding-boxes - obb - rotation-aware-tracking - object-detection - tracking - mot - mmot - arxiv:2510.12565 size_categories: - 100K This repository is a Hugging Face packaging of MMOT. Please cite the original MMOT paper and follow the official dataset license and redistribution terms. ## Dataset Details ### Dataset Description - **Dataset name:** MMOT - **Full name:** MMOT: The First Challenging Benchmark for Drone-based Multispectral Multi-Object Tracking - **Task:** Multispectral multi-object tracking, drone-based MOT, oriented-object tracking, object detection - **Domain:** UAV / aerial video, urban and rural scenes, transportation scenes, public spaces, playgrounds, and sports courts - **Data type:** 8-band multispectral frame arrays with MOT-style oriented annotations - **Frame format:** `.npy` multispectral arrays - **Annotation format:** `.txt` MOT-style files with oriented bounding-box corner coordinates - **Official code / project:** [Annzstbl/MMOT](https://github.com/Annzstbl/MMOT) - **Paper:** [MMOT: The First Challenging Benchmark for Drone-based Multispectral Multi-Object Tracking](https://arxiv.org/abs/2510.12565) - **Hugging Face paper page:** [huggingface.co/papers/2510.12565](https://huggingface.co/papers/2510.12565) MMOT was introduced to address the lack of dedicated UAV multispectral MOT datasets. The benchmark emphasizes conditions where RGB-only spatial appearance cues can be weak, including extreme small targets, dense instances, severe occlusion, blur, cluttered backgrounds, fast object motion, irregular UAV motion, and in-plane rotation. ### Supported Tasks This dataset can be used for: - Multispectral multi-object tracking - Drone-based / UAV multi-object tracking - Multi-class tracking in aerial scenes - Oriented bounding-box detection and tracking - Rotation-aware tracking-by-detection - Rotation-aware tracking-by-query - Multispectral feature learning - Tracking association under small-object and high-density conditions - MOT benchmark conversion and evaluation pipelines ## Dataset Structure The Hugging Face release packages video sequences as `.tar` archives. Each archive contains multispectral `.npy` frames and per-frame `.txt` annotations. A typical Hugging Face layout is: ```text MMOT_DATASET/ ├── train/ │ ├── data30-8.tar │ ├── data23-2.tar │ └── ... ├── test/ │ ├── dataXX-X.tar │ └── ... └── README.md ``` After extraction, each sequence archive contains frame-level files such as: ```text data30-8/ ├── 000001.npy ├── 000001.txt ├── 000002.npy ├── 000002.txt └── ... ``` The official repository includes a helper script that converts the Hugging Face archive layout into the standard MMOT directory layout: ```bash python dataset/huggingface_tar_to_standard.py --root /path/to/MMOT_DATASET ``` After conversion, the expected standard layout is: ```text MMOT_DATASET/ ├── train/ │ ├── npy/ │ │ ├── data23-2/ │ │ │ ├── 000001.npy │ │ │ ├── 000002.npy │ │ │ └── ... │ │ └── ... │ └── mot/ │ ├── data23-2.txt │ ├── data23-3.txt │ └── ... └── test/ ├── npy/ │ └── ... └── mot/ └── ... ``` Each standard-layout sequence contains: ```text npy/SEQ_NAME/ # Multispectral frames as .npy arrays mot/SEQ_NAME.txt # Merged sequence-level MOT-style annotations ``` ## Splits MMOT is split into training and test sets. The paper reports that the split avoids overlap of geographic locations and specific scene instances across subsets. | Split | Sequences | Frames | Identity tracks | Oriented boxes | |---|---:|---:|---:|---:| | `train` | 75 | 8,372 | 6,101 | ~292K | | `test` | 50 | 5,446 | 4,527 | ~196K | | **Total** | **125** | **13,818** | **10,628** | **~488K** | ## Dataset Statistics The MMOT paper and official repository report the following high-level statistics: | Statistic | Value | |---|---:| | Video sequences | 125 | | Frames | 13.8K | | Total duration | 2,767 seconds | | Oriented bounding-box annotations | ~488.8K | | Average annotations per frame | 35.2 | | Object categories | 8 | | Spectral channels | 8 | | Spatial resolution | 1200 × 900 | | Flight altitude range | 80–200 m | | Maximum objects per frame | 155 | | Average objects within 300 px radius | 19.4 | ### Categories MMOT contains three superclasses and eight fine-grained object categories: | Superclass | Categories | |---|---| | `HUMAN` | pedestrian | | `VEHICLE` | car, van, truck, bus | | `BIKE` / `BICYCLE` | bike, tricycle, awning-bike | The official TrackEval configuration uses the following class IDs: | Class ID | Class name | |---:|---| | 0 | car | | 1 | bike | | 2 | pedestrian | | 3 | van | | 4 | truck | | 5 | bus | | 6 | tricycle | | 7 | awning-bike | ## Annotation Format MMOT uses MOT-style text files with oriented bounding-box coordinates. ### Sequence-Level MOT-OBB Format After conversion to the standard layout, annotations are stored as: ```text train/mot/SEQ_NAME.txt test/mot/SEQ_NAME.txt ``` Each row is comma-separated and follows the format used by the official MMOT tools: ```text , , , , , , , , , , , , ``` Field descriptions: | Field | Description | |---|---| | `frame` | Frame index, starting from 1 | | `id` | Object identity ID within the sequence | | `x0, y0, ..., x3, y3` | Four OBB corner points in image coordinates | | `score_or_flag` | Score, confidence, or validity field depending on ground-truth vs. prediction files | | `class` | Category ID, using the class mapping above | | `truncation_or_extra_flag` | Extra annotation flag; commonly used for truncation-related metadata | The eight corner-coordinate fields represent the oriented bounding box directly. For code that expects rotated rectangles such as `(cx, cy, w, h, angle)`, convert these corner points using your preferred OBB geometry library. ### YOLO-OBB Conversion Format The official repository includes a `mot2yolo_obb.py` converter. Its YOLO-OBB output rows use normalized corner coordinates: ```text ``` where all coordinates are normalized by the frame width and height. ## Usage ### Download from Hugging Face ```python from huggingface_hub import snapshot_download repo_dir = snapshot_download( repo_id="YOUR_USERNAME_OR_ORG/MMOT", repo_type="dataset", ) print(repo_dir) ``` Replace `YOUR_USERNAME_OR_ORG/MMOT` with the actual Hugging Face dataset repository ID. ### Convert Hugging Face Archives to the Standard Layout ```bash python dataset/huggingface_tar_to_standard.py --root /path/to/MMOT_DATASET ``` This extracts each sequence archive into `/npy//` and merges per-frame `.txt` files into `/mot/.txt`. ### Example: Load a Multispectral Frame ```python from pathlib import Path import numpy as np root = Path("MMOT_DATASET") seq = "data23-2" frame_path = root / "train" / "npy" / seq / "000001.npy" frame = np.load(frame_path) print(frame.shape) # Usually H x W x 8, depending on the saved array layout ``` ### Example: Read MOT-OBB Annotations with Python ```python from pathlib import Path import pandas as pd root = Path("MMOT_DATASET") seq = "data23-2" ann_path = root / "train" / "mot" / f"{seq}.txt" columns = [ "frame", "id", "x0", "y0", "x1", "y1", "x2", "y2", "x3", "y3", "score_or_flag", "class", "truncation_or_extra_flag", ] ann = pd.read_csv(ann_path, header=None, names=columns) print(ann.head()) print(ann["id"].nunique(), "unique object tracks") ``` ### Example: Get Annotations for One Frame ```python frame_idx = 1 frame_ann = ann[ann["frame"] == frame_idx] corners = frame_ann[["x0", "y0", "x1", "y1", "x2", "y2", "x3", "y3"]].to_numpy() classes = frame_ann["class"].astype(int).to_numpy() ids = frame_ann["id"].astype(int).to_numpy() print(corners.shape, classes.shape, ids.shape) ``` ### Example: Create Pseudo-RGB from Multispectral Data The MMOT paper evaluates RGB baselines by synthesizing pseudo-RGB images from MSI bands 5, 3, and 2. Adjust indexing according to your array convention, because Python uses zero-based indexing. ```python import numpy as np frame = np.load("MMOT_DATASET/train/npy/data23-2/000001.npy") # If frame is H x W x 8 and the paper's band numbers are treated as 1-based: pseudo_rgb = frame[:, :, [4, 2, 1]] # Optional normalization for visualization only pseudo_rgb = pseudo_rgb.astype("float32") pseudo_rgb = (pseudo_rgb - pseudo_rgb.min()) / (pseudo_rgb.max() - pseudo_rgb.min() + 1e-6) ``` ## Evaluation MMOT is evaluated with MOT-style metrics, including: - HOTA - MOTA - IDF1 - DetA - AssA - CLEAR MOT metrics The official repository integrates TrackEval for MMOT evaluation. The paper reports two aggregation protocols for the multi-class setting: - **Class-averaged evaluation:** average metrics across classes. - **Detection-averaged evaluation:** aggregate detections across classes before computing metrics. The paper evaluates both RGB and MSI input settings: - **RGB setting:** pseudo-RGB images are synthesized by selecting bands 5, 3, and 2 from the MSI cube. - **MSI setting:** all 8 spectral channels are used. For official benchmark reporting, use the MMOT authors' evaluation code and protocol. For local experiments: - Keep training and test sequences separate. - Do not tune hyperparameters on test-set ground truth. - Report whether the tracker uses RGB, pseudo-RGB, or full MSI input. - Report whether OBBs are evaluated as corner polygons, rotated rectangles, or converted axis-aligned boxes. - Use the expected sequence-local identity IDs and per-sequence output files. ## Dataset Creation ### Source Data MMOT was collected using a UAV equipped with a downward-facing multispectral camera. The camera captures 8 spectral bands spanning the visible to near-infrared range. The data were collected at dynamic flight altitudes from 80 m to 200 m, under conditions including clear skies, cloudy days, and dense fog. The benchmark covers diverse scenes such as urban streets, rural fields, traffic intersections, transit hubs, playgrounds, and sports courts. Frames were registered across spectral channels and cropped to 1200 × 900 pixels. ### Annotation Guidelines MMOT uses oriented bounding boxes because arbitrary object orientation is common in aerial views. The annotation process assigns a unique identity to each object across frames and uses a multi-stage annotation pipeline. The paper describes the following annotation principles: - Exhaustively annotate instances in the predefined categories. - Use spectral channels to assist when targets are difficult to distinguish in pseudo-color views. - Validate ambiguous targets using temporal context across the full sequence. - Label the full object extent, including under occlusion, truncation, or motion blur. - Maintain consistent identity IDs over time. The annotation workflow includes initial box placement, box refinement, identity assignment, identity correction, and expert-level cross-validation. ## Intended Use MMOT is intended for academic research and benchmarking in: - Multispectral multi-object tracking - Drone-based MOT - Oriented and rotation-aware tracking - Multi-class aerial tracking - Robust association under small-object, dense, occluded, or fast-motion conditions - Multispectral representation learning ## Out-of-Scope Use This dataset should not be used for: - Surveillance deployments targeting private individuals without appropriate legal basis and safeguards - Biometric identification or person identification outside the dataset's research context - Commercial use that violates the dataset license - Redistribution, modification, or derivative release that violates the original MMOT terms - Claims of official benchmark performance without following the official evaluation protocol ## Limitations and Biases MMOT focuses on UAV-based multispectral tracking and may not generalize to ground-level videos, fixed surveillance cameras, non-UAV sensors, different camera spectral responses, or domains outside the covered scene types. Because the data are collected from aerial views, many objects are small and appearance cues can be weak. Models trained only on MMOT may overfit to the dataset's specific flight altitudes, object categories, geographic environments, weather conditions, spectral bands, and annotation conventions. The dataset is valuable for research on multispectral tracking, but high-quality OBB annotation is labor-intensive. The paper identifies scalable annotation and unsupervised learning as future directions. ## License and Terms The MMOT dataset is released under the **Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0)** according to the official repository and Hugging Face listing. The dataset is intended for academic research. Users must attribute the original source and should not modify or redistribute the dataset without permission. Please consult the official MMOT repository, dataset page, and license before redistributing, modifying, or using the dataset commercially. Code in the official repository may contain multiple submodules with their own licenses. Check the license file or terms in each subdirectory before reusing code. ## Citation If you use MMOT, please cite the original paper: ```bibtex @inproceedings{li2025mmot, title = {MMOT: The First Challenging Benchmark for Drone-based Multispectral Multi-Object Tracking}, author = {Li, Tianhao and Xu, Tingfa and Wang, Ying and Qin, Haolin and Lin, Xu and Li, Jianan}, booktitle = {NeurIPS 2025 Datasets and Benchmarks Track}, year = {2025}, url = {https://arxiv.org/abs/2510.12565} } ``` ## Acknowledgements MMOT was created by researchers from Beijing Institute of Technology and Beijing Institute of Technology Chongqing Innovation Center. Please refer to the original paper and official repository for complete author information, benchmark details, code, pretrained models, and updates.