data_juicer.ops.mapper.video_human_tracks_extraction_mapper module

class data_juicer.ops.mapper.video_human_tracks_extraction_mapper.VideoHumanTracksExtractionMapper(*args, **kwargs)

Bases: Mapper

Extract face and human bounding box tracks from videos.

This operator performs multi-stage processing including scene detection, face detection (S3FD), face tracking, and human detection (YOLOv8). It eventually generates synchronized face and human tracks and saves the bbox sequences into pickle files.

Source: This operator is a part of HumanVBench (CVPR 2026).

__init__(face_track_bbox_path: str = './HumanVBenchRecipe/dj_human_track', YOLOv8_human_model_path: str = './thirdparty/humanvbench_models/YOLOv8_human/weights/best.pt', face_detect_S3FD_model_path: str = './thirdparty/humanvbench_models/Light-ASD/model/faceDetector/s3fd/sfd_face.pth', tag_field_name_human_track_path: str = 'human_track_data_path', tag_field_name_people_num: str = 'number_people_in_video', *args, **kwargs)

Initialization method.

get_face_and_human_tracks(video_array, track, human_detection_pipeline)

process_single(sample, rank=None)

For sample level, sample –> sample

Parameters:

sample – sample to process

Returns:

processed sample