How sejong-rcv/insanet works
Project type and purpose
• This is a specialized machine‐learning research codebase (not a web app or microservice). It implements INSANet, a PyTorch model for multispectral pedestrian detection, with utilities for data download/prep, training, inference, evaluation, and visualization.
File structure
• Top-level folders:
• data: download scripts & raw dataset layout
• docker: Dockerfile + Makefile + Python requirements
• evaluation: COCO conversions, benchmark scripts, state-of-art results
• pretrained: scripts to fetch pretrained weights
• src: core implementation (config, datasets, model, train_eval.py, inference.py), plus subfolders:
– script: shell wrappers for train/infer
– utils: helper code (transforms, evaluation scripts)
– imageSets & annotations: dataset index files and JSON
– results: stored detection outputs
• fig: architecture & paper figures
• environment*.yaml: Conda environment specs
README for additional insights
• Describes the high-level model architecture (INtra/INter spectral attention)
• Describes Docker vs. Conda setup
• Presents demo, train & inference procedures, evaluation protocols
Main components and interactions
a. Data Layer
• External dataset sources (KAIST, LLVIP)
• data/download_*.sh/.py → populates data/kaist-rgbt
• data loader (src/datasets.py + src/imageSets + JSON annotations)
b. Environment & Deployment
• Docker container (docker/Dockerfile, Makefile)
• Conda env (environment.yaml, environment_cu113.yaml)
c. Core Model & Training
• Model definition (src/model.py, src/insa.py)
• Configuration (src/config.py)
• Training script (src/train_eval.py + src/script/train_eval.sh)
• Intra-/Inter-Spectral Attention blocks—key architectural pattern
d. Inference
• Inference entrypoint (src/inference.py + src/script/inference.sh)
• Loads pretrained weights (pretrained/download_pretrained.py)
e. Evaluation & Visualization
• COCO conversion & evaluation (src/utils/coco*.py, evaluation/coco*.py)
• Scene-specific evaluation (src/utils/evaluation_scene*)
• Benchmark plotting (evaluation/evaluation_script.py)
• Visualization tool (vis.py)
f. Outputs & Artifacts
• Trained models (checkpoints)
• Detection results (src/results)
• Figures (fig/, evaluation/*.jpg)
Interactions:
Data Layer → DataLoader → Model → Training/Infinference → Results → Evaluation → Figures
Architectural patterns & design principles
• Pipeline pattern: clearly separated stages (data download → preprocessing → training → inference → evaluation)
• Modular organization: config, dataset, model, utils, scripts
• Containerized reproducibility: Docker + Conda
• Attention-based fusion: INSANet blocks in model graph