How ishitsuka-hikaru/safeuav works
Project type and purpose
• SafeUAV is a research-driven computer-vision/ML toolkit for UAV depth estimation and safe-landing area prediction.
• It provides training, testing, and inference scripts around deep segmentation and regression models (UNet variants and DeepLabv3+).
Main components of the system
a. Data Layer
– Synthetic/real dataset stored as HDF5 files (.h5)
– Optional label-export script (main_export_new_labels.py)
b. Model Definitions
– unet_tiny_sum.py, unet_classic.py, deeplabv3plus_xception.py, etc.
– Loss functions module (loss.py)
c. Training Pipeline
– Orchestrator: main.py
– Dataset loader & sampler (test_dataset.py)
– Optimizer, scheduler, checkpoint logic
– Plotting metrics & outputs (plotter.py)
d. Inference Pipelines
– Video inference: main_inference_video.py
– Camera/real-time inference: main_inference_camera.py
e. Deployment & Environment
– Docker build (docker/Dockerfile, docker/requirements.txt)
– Jetson installation prerequisites (packages.txt)
– External neural-wrappers dependency
Relationships and data flows
• Data Layer → Training Pipeline
– HDF5 dataset is read by main.py
– Model definitions and loss.py are injected
– Training produces model weights (“.pkl” checkpoints)
• Training Pipeline → Plotter
– Periodic evaluation outputs curves/images via plotter.py
• Checkpoint Storage → Inference Pipelines
– main_inference_video.py / main_inference_camera.py load weights
– Input (video file or camera stream) → model → post-processing → output video/frames
• Entire workflow can be containerized in Docker or deployed on Jetson
Architectural patterns and design principles
– Modular separation: model definitions vs. training logic vs. inference logic
– Pipeline orchestration: single-entry scripts for each stage (train/test/infer)
– Configuration-driven experiments: command-line flags control task (classification vs. regression), model, hyperparameters
– Reproducibility via Docker or environment scripts
Elements to include for maximum clarity
– Versioned artifacts (e.g., model-checkpoint.pkl)
– Hardware boundary: “GPU / Jetson”
– Container boundary labeled “Docker Image” with exposed volumes