Computer Vision Toolkit

Computer Vision Toolkit

End-to-end computer vision framework for image classification, object detection, and semantic segmentation — with real mAP/IoU math, PyTorch training loops, data augmentation, and model export.

Train, evaluate, and deploy CV models using a single coherent codebase. Every evaluation metric is implemented with real math (no black-box library calls), every training loop supports mixed precision and gradient accumulation, and the augmentation pipeline composes transforms the way you'd compose functions.

What You Get

• Image classification — build classification pipelines with pre-trained ResNet/EfficientNet/ViT backbones, custom heads, and label management
• Object detection — anchor-based detection with configurable anchor scales, multi-scale feature maps, and NMS post-processing
• Semantic segmentation — UNet and FPN decoder architectures on any backbone, with per-pixel loss computation and boundary refinement
• Data augmentation library — composable transforms (flip, rotate, crop, color jitter, blur, grayscale) with Compose, OneOf, and per-transform probability control
• Dataset loaders — ImageFolderDataset for classification, DetectionDataset for COCO/VOC/YOLO-format annotations, with lazy loading and caching
• Training loops — full PyTorch training with mixed precision (AMP), gradient accumulation, cosine/step/plateau LR scheduling, warmup, and automatic checkpointing
• Evaluation framework — real implementations of IoU, mAP@0.5, mAP@[.5:.95] (COCO-style), precision-recall curves, confusion matrices, and segmentation mIoU — all using stdlib math
• Inference engine — batch inference, confidence thresholding, NMS, and ONNX export
• Model architectures — backbone library (ResNet, EfficientNet, MobileNet, ViT) and task-specific heads (classification, detection, segmentation)

File Tree

computer-vision-toolkit/
├── README.md
├── LICENSE
├── requirements.txt
├── src/
│   └── cv_toolkit/
│       ├── __init__.py              # Package exports
│       ├── classification.py        # Classification pipeline and models
│       ├── detection.py             # Object detection with anchor configs
│       ├── segmentation.py          # Semantic/instance segmentation
│       ├── augmentation.py          # Composable augmentation transforms
│       ├── datasets.py              # Dataset loaders (ImageFolder, COCO, VOC)
│       ├── training.py              # Training loop with AMP and scheduling
│       ├── evaluation.py            # mAP, IoU, classification metrics (real math)
│       ├── inference.py             # Batch inference and model export
│       └── models/
│           ├── __init__.py          # Model registry
│           ├── backbones.py         # ResNet, EfficientNet, MobileNet, ViT
│           └── heads.py             # Classification, detection, segmentation heads
├── configs/
│   ├── training_config.yaml         # Training hyperparameter presets
│   └── augmentation_config.yaml     # Augmentation pipeline configurations
├── tests/
│   ├── test_evaluation.py           # mAP, IoU, classification metric tests
│   ├── test_augmentation.py         # Augmentation pipeline tests
│   └── test_detection.py            # Detection model and anchor tests
└── guides/
    ├── training-guide.md            # Training workflow walkthrough
    └── evaluation-guide.md          # Metrics explained with decision guidance

Quick Start

1. Install Dependencies

pip install -r requirements.txt

*... continues with setup instructions, usage examples, and more.*