How yangtaowang95/tokencut works
TokenCut is a research application for discovering objects and salient regions from self-supervised vision-transformer features, with a separate weakly supervised detection workflow. The primary path takes an image or dataset through feature extraction, normalized-cut segmentation, and bounding-box or mask output; evaluation and visualization report or render results. The README also documents weakly supervised training and evaluation. Some connections come from the README or SOURCE INDEX rather than complete excerpts, so unsampled implementation details are not expanded into additional flows.
Main entry points: get_saliency.py Saliency Runner · main_tokencut.py TokenCut Runner · eval.py Detection Evaluation · main.py Detection Training
Components 23
Object Discovery
- TokenCut Runner
main_tokencut.py - Normalized Cut
object_discovery.py - Discovery Visualizations
visualizations.py
Saliency Detection
- Saliency Runner
unsupervised_saliency_detection/get_saliency.py - Saliency Features
unsupervised_saliency_detection/dino.py - Saliency Cut
unsupervised_saliency_detection/object_discovery.py - Bilateral Solver
unsupervised_saliency_detection/bilateral_solver.py - Saliency Metrics
unsupervised_saliency_detection/metric.py - Saliency Results
unsupervised_saliency_detection/get_saliency.py
Vision Models
- Model Loader
networks.py - DINO Vision Transformer
dino/vision_transformer.py
Data Inputs
- Discovery Datasets
datasets.py
Weak Detection
- Detection Training
weakly_supvervised_detection/main.py - Detection Datasets
weakly_supvervised_detection/datasets.py - Training Samplers
weakly_supvervised_detection/samplers.py - Detection Transformer
weakly_supvervised_detection/vision_transformer.py - Linear Fine-tuning
weakly_supvervised_detection/finetune.py - Detection Evaluation
weakly_supvervised_detection/eval.py - Detection Metrics
weakly_supvervised_detection/metrics.py - Evaluation Meters
weakly_supvervised_detection/meters.py - Detection Visualizations
weakly_supvervised_detection/visualization.py
Other components
- Researcher
- Input Images
Connections 24
- Researcher → TokenCut Runner: runs discovery
README.md - Input Images → TokenCut Runner: supplies image
README.md - TokenCut Runner → Discovery Datasets: loads dataset
main_tokencut.py - TokenCut Runner → Model Loader: loads model
main_tokencut.py - Model Loader → DINO Vision Transformer: constructs transformer
networks.py - TokenCut Runner → Normalized Cut: computes cut
main_tokencut.py - TokenCut Runner → Discovery Visualizations: renders results
main_tokencut.py - Researcher → Saliency Runner: runs saliency
README.md - Input Images → Saliency Runner: supplies image
README.md - Saliency Runner → Saliency Features: extracts features
unsupervised_saliency_detection/get_saliency.py - Saliency Runner → Saliency Cut: computes cut
unsupervised_saliency_detection/get_saliency.py - Saliency Runner → Bilateral Solver: refines mask
unsupervised_saliency_detection/get_saliency.py - Saliency Runner → Saliency Metrics: evaluates mask
unsupervised_saliency_detection/get_saliency.py - Saliency Runner → Saliency Results: writes results
unsupervised_saliency_detection/get_saliency.py - Researcher → Detection Training: trains detector
README.md - Detection Training → Detection Datasets: builds datasets
weakly_supvervised_detection/main.py - Detection Training → Training Samplers: selects samples
weakly_supvervised_detection/main.py - Detection Training → Detection Transformer: uses transformer
weakly_supvervised_detection/main.py - Researcher → Linear Fine-tuning: runs fine-tuning
README.md - Researcher → Detection Evaluation: runs evaluation
README.md - Detection Evaluation → Detection Datasets: loads dataset
weakly_supvervised_detection/eval.py - Detection Evaluation → Detection Metrics: computes metrics
weakly_supvervised_detection/eval.py - Detection Evaluation → Evaluation Meters: tracks results
weakly_supvervised_detection/eval.py - Detection Evaluation → Detection Visualizations: renders detections
weakly_supvervised_detection/eval.py