How madhu-basavanna/computer_vision works
The repository is an experimental vision pipeline for learning self-supervised image features and using them for semantic or class-conditioned segmentation, with a separate nearest-neighbor analysis workflow. Prepared image crops or annotated COCO data enter dataset readers and training scripts; the resulting models produce pixel masks and segmentation metrics. The README documents the principal pipeline and capabilities. Some subsystem links rely on that documentation because the relevant source was not excerpted; those limits are reflected in the evidence citations.
Main entry points: dt_multiclass_ss.py Multiclass training · dt_single_ss.py Class-conditioned training · pretrain.py DINO training · pretraining_backbone.py ResNet feature backbone
Components 18
Data preparation
- Crop reader
data/pretraining.py - Annotation readers
data/segmentation.py - Crop augmentations
data/data_augmentation.py - Image transforms
data/transforms.py
Representation learning
- DINO training
pretrain.py - ResNet feature backbone
models/pretraining_backbone.py
Segmentation
- Multiclass training
dt_multiclass_ss.py - Class-conditioned training
dt_single_ss.py - Multiclass segmentator
models/second_segmentation.py - Attention segmentator
models/att_segmentation.py - Class-conditioned attention
models/attention_layer.py - DeepLab decoder
models/deeplab_decoder.py
Analysis and support
- Nearest-neighbor analysis
nearest_neighbors.py - Masks and mIoU
- Checkpoint loading
utils/weights.py - Training meters
utils/meters/averagevaluemeter.py - Meter base
utils/meters/meter.py
Other components
- Prepared datasets
Connections 26
- Prepared datasets → Crop reader: provides crops
README.md - Prepared datasets → Annotation readers: provides annotations
README.md - DINO training → Crop reader: loads images
pretrain.py - DINO training → Crop augmentations: applies crops
pretrain.py - DINO training → ResNet feature backbone: trains networks
pretrain.py - DINO training → Masks and mIoU: exports analysis
README.md - Nearest-neighbor analysis → ResNet feature backbone: extracts features
README.md - Nearest-neighbor analysis → Crop reader: reads crops
README.md - Multiclass training → Annotation readers: loads masks
dt_multiclass_ss.py - Multiclass training → Image transforms: prepares inputs
dt_multiclass_ss.py - Multiclass training → ResNet feature backbone: uses features
dt_multiclass_ss.py - Multiclass training → Multiclass segmentator: trains model
dt_multiclass_ss.py - Multiclass segmentator → DeepLab decoder: decodes features
models/second_segmentation.py - Multiclass training → Checkpoint loading: loads checkpoint
dt_multiclass_ss.py - Multiclass training → Training meters: tracks training
dt_multiclass_ss.py - Multiclass training → Masks and mIoU: evaluates masks
README.md - Class-conditioned training → Annotation readers: loads masks
dt_single_ss.py - Class-conditioned training → Image transforms: prepares inputs
dt_single_ss.py - Class-conditioned training → ResNet feature backbone: uses features
dt_single_ss.py - Class-conditioned training → Attention segmentator: trains model
dt_single_ss.py - Attention segmentator → Class-conditioned attention: applies attention
models/att_segmentation.py - Attention segmentator → DeepLab decoder: decodes features
models/att_segmentation.py - Class-conditioned training → Checkpoint loading: loads checkpoint
dt_single_ss.py - Class-conditioned training → Training meters: tracks training
dt_single_ss.py - Class-conditioned training → Masks and mIoU: evaluates masks
README.md - Training meters → Meter base: extends
utils/meters/averagevaluemeter.py