How riips/aic24_track1_yachiyo_riips works
Project Type and Purpose
• This is an offline, multi-stage computer vision pipeline for people tracking across multiple cameras (the AI City Challenge Track 1 submission).
• It ingests raw video files, extracts frames, detects persons, computes embeddings (re‐identification features), estimates poses, performs single‐ and multi‐camera tracking, and outputs a unified tracking result for submission.
High-Level Components
a. Input Layer
• Raw video files organized by scene/camera (e.g., Original/scene_041/camera_0361/video.mp4)
b. Preprocessing
• Frame Extraction (scripts/extract_frame.sh → tools/extract_frame.py)
c. Detector Module
• Person detection via BoT-SORT (detector/aic24_get_detection.py invoked by scripts/detection.sh)
d. Embedder Module
• Feature extraction with deep‐person‐reid (embedder/aic24_extract.py via scripts/embedding.sh)
• Produces appearance embeddings per detected person
e. Pose Estimation Module
• Top-down pose estimation with mmpose (poser/top_down_video_demo_with_track_file.py via scripts/pose.sh)
• Produces keypoint annotations per person crop
f. Tracking Module
• Single‐camera tracking (mcpt.py)
• Multi‐camera tracking (scpt.py, run.py orchestrating clustering)
• Configurable per-scene parameters (tracking/config/*.py, JSON)
• Entry point: tracking/infer.py or scripts/tracking.sh
g. Submission Generator
• tools/generate_submission.py combines all scene outputs into a submission file
Architectural Patterns & Principles
• Modular, pipeline‐oriented design: each stage is independent, testable, and replaceable.
• Clear separation of concerns: detection, embedding, pose, tracking, and I/O are distinct.
• Configuration‐driven tracking: per-scene parameters and camera mappings live in config files.
• Leverages external frameworks: BoT-SORT, torchreid, and mmpose, isolating wrappers in detector/, embedder/, and poser/.