OBBPose dataset for direction-aware detection, pose tracking, and fish behavior analysis

Description

This dataset supports the study "A Direction-Aware Detection and Pose-Tracking Framework for Quantifying Micropterus salmoides Swimming Behavior." It contains 861 paired images and oriented-bounding-box/pose annotations, including 601 training, 173 validation, and 87 test samples. The package also contains three tracking-evaluation videos with frame-level XML annotations, selected model weights for seeds 0, 111, and 222, available training records, behavioral-validation time series, dataset manifests, and machine-readable values reported in the manuscript. Some comparison rows in manuscript Tables 1 and 2 do not have complete per-seed raw records in the source project. Their reported summary values are preserved in the paper_tables directory with provenance information. The accompanying software is available at: https://github.com/XuRanHao/OBBPose Archived software release: https://doi.org/10.5281/zenodo.21802941

Steps to reproduce

1. Download OBBPose-public-data-v1.0.0.zip and verify its SHA-256 checksum using SHA256SUMS.txt. 2. Extract the archive. Dataset structure, field definitions, and split information are documented in README.md, DATA_DICTIONARY.md, split_manifest.csv, and tracking_sequence_manifest.csv. 3. Obtain the accompanying software from: https://github.com/XuRanHao/OBBPose 4. Create the reference environment: conda env create -f environment.yml conda activate obbpose 5. Place OBBPose-public-code and OBBPose-public-data as sibling directories. 6. Reproduce the available training configurations over seeds 0, 111, and 222 with: python run_grid.py --data ../OBBPose-public-data/detection_dataset/dataset.yaml 7. Tracking evaluation and behavioral quantification commands are provided in the software README. Released training records, weights, behavioral results, and manuscript table values are included in the archive. Some comparison rows in manuscript Tables 1 and 2 do not have complete per-seed raw records in the source project. These rows are retained as reported summary values in paper_tables with provenance information.

Institutions

• Shanghai Ocean University

  Shanghai, Shanghai

Categories

Related Links

Licence