gangweix/fast-acvnet

How gangweix/fast-acvnet works

Fast-ACVNet is a stereo-matching research application: it trains and evaluates disparity-estimation models on Scene Flow and KITTI, and provides benchmark-output and Middlebury evaluation paths. The main training scripts coordinate dataset selection, model execution, loss and metrics, and experiment logging; the README documents staged attention-only then complete-network training. Dataset loaders also cover ETH3D and KITTI raw data. Some subsystem paths and links come from the README or source index rather than inspected implementation, so uncertain runtime connections are omitted.

Main entry points: main_kitti.py KITTI training · main_sceneflow.py Scene Flow training · test_mid.py Middlebury evaluation · Fast_ACV_plus.py Fast-ACVNet+

Components 18

Training workflow

Scene Flow training main_sceneflow.py
KITTI training main_kitti.py

Stereo data

Scene Flow loader datasets/sceneflow_dataset.py
KITTI loader datasets/kitti_dataset_1215.py
Image and PFM I/O datasets/data_io.py
Flow transforms datasets/flow_transforms.py
Other datasets datasets

Disparity models

Fast-ACVNet models/Fast_ACV.py
Fast-ACVNet+ models/Fast_ACV_plus.py
Model building blocks models/submodule.py
Training loss models/loss.py

Evaluation and output

Stereo metrics utils/metrics.py
Experiment utilities utils/experiment.py
Visualization utils/visualization.py
Runtime utilities utils/misc.py
Middlebury evaluation test_mid.py
Benchmark disparity output save_disp.py

Other components

Researcher

Connections 22

Researcher → Scene Flow training: starts training README.md
Researcher → KITTI training: starts training README.md
Scene Flow training → Scene Flow loader: loads batches main_sceneflow.py
KITTI training → KITTI loader: loads batches main_kitti.py
KITTI loader → Image and PFM I/O: reads images datasets/kitti_dataset_1215.py
Scene Flow loader → Image and PFM I/O: reads images datasets/sceneflow_dataset.py
Middlebury evaluation → Flow transforms: augments samples
Middlebury evaluation → Image and PFM I/O: uses transforms
Middlebury evaluation → Other datasets: evaluates
Scene Flow training → Fast-ACVNet: runs model main_sceneflow.py
KITTI training → Fast-ACVNet+: runs model main_kitti.py
Fast-ACVNet → Model building blocks: uses blocks models/Fast_ACV.py
Fast-ACVNet+ → Model building blocks: uses blocks models/Fast_ACV_plus.py
Scene Flow training → Training loss: computes loss
KITTI training → Training loss: computes loss main_kitti.py
KITTI training → Stereo metrics: computes metrics main_kitti.py
KITTI training → Experiment utilities: logs results main_kitti.py
Experiment utilities → Visualization: renders images
KITTI training → Runtime utilities: initializes runtime
Researcher → Benchmark disparity output: requests benchmark output README.md
Researcher → Middlebury evaluation: evaluates generalization README.md
Benchmark disparity output → Fast-ACVNet+: loads predictions

Drawn by GitDiagram on November 9, 2026 from the repository's file tree, README and source files. Read as Markdown ·