How lijia7/sddr works
Project type and purpose
• This is a research‐code Python pipeline for “Self‐Distilled Depth Refinement with Noisy Poisson Fusion,” implemented as a CLI/inference tool (not a web app).
• It takes stereo/rgb inputs, runs one‐ or two‐stage depth estimation (LeRes, MiDaS, SGR submodels), fuses/refines them, and writes out depth maps.
File structure
• Top-level:
• demo.py, previous_evaluate_mid21.py – entry‐point scripts (CLI interface, evaluation).
• requirements.txt – Python dependencies.
• checkpoints/ – pretrained PyTorch model weights.
• demo_input/, demo_output/ – folders for test data and results.
• utils/ – project‐specific helper modules (backbone definitions, fusion functions, loss, model wrappers).
• thirdrepo/ – three external repositories integrated as subpackages:
– LeRes/ (ResNet‐based depth net)
– MiDaS/ (MiDaS base model and transforms)
– SGR/ (another depth network + custom synchronized‐batchnorm C/CUDA extension)
• assets/ – static assets (e.g., NeurIPS logo).
README for additional insights
• The README confirms the CLI usage (one-stage vs two-stage), weight arguments, input/output paths, and evaluation script.
• It points out core steps: load base depth predictions (high/low resolution), refine via Poisson fusion and self‐distillation.
Elements
CLI → Data Loader → Base Depth Nets → Fusion Engine → Output Writer → Evaluation
– Blue for code modules (demo.py, utils, thirdrepo)
– Green for data artifacts (demo_input, checkpoints, demo_output)
– Orange for external/C++ extension (SGR syncbn)
Final advice
• Core algorithmic stages (base net inference, Poisson fusion, self-distillation).