A repository inherited from COCO-api, for person detection/pose estimation benchmark.
This is an unofficial re-implementation of COCO-api.
Why do we need another version of COCO-api?: For the sake of solving the limitation of COCO-api including:
- The greedy-search matching, used in COCO api seems to cause problem in matching process;
- Maximum detection is limited to 20 for keypoints/100 for detection validation;
- Parameters are casted implicitly in evaluation api, very hard to re-implement to adapt to different datasets;
- No visualization function;
- Add few tools maybe used in benchmark;
Main functions: Basically, hie project inherits from COCO-api, it contains every function included in COCO and also the original evaluation method in COCO.
tools: Some basic simple functions often used in objection detction/pose estimation benchmark.
0.0.0: The original version was derived from PoseBenchmark project;
0.1.0: Initialization;
0.1.1: Replace original COCO-api: ComputeIOU with HIE-api: oks;
0.1.2: Add test functions;
0.1.3: Fix time and copy bug;
0.1.4: Fix PYTHONVERSION bug;
0.1.5: Add HIE.load_res() test;
0.1.6: Fake commit;
0.1.7: Add img_files and a scalable viz function for HIE and HIEval;
0.1.8:
- Add show_bbox only for viz_ann funciton;
- Fix HIEval viz bug: always loading hie dataset images;
0.1.9:
- Fix HIEval viz bug: no show_bbox_only;
- Add evaluation for tracking;