majiayu000 Computer Vision
Build computer vision systems for image classification, object detection, and visual analysis.
Overview
Computer vision enables machines to interpret and understand visual information. This skill covers CNNs, object detection, segmentation, and modern vision architectures.
Quick Reference
| Scenario | Recommended Approach | Reference File |
|---|---|---|
| Image classification | CNNs (ResNet, EfficientNet, ViT) | /references/classification.md |
| Object detection and localization | YOLO, Faster R-CNN, RetinaNet | /references/object-detection.md |
| Pixel-level segmentation | U-Net, Mask R-CNN, DeepLab | /references/segmentation.md |
Core Principles
- Convolutional Layers - Extract spatial features hierarchically
- Transfer Learning - Use pre-trained models on ImageNet
- Data Augmentation - Increase robustness through transformations
- Multi-Scale Processing - Detect objects at different sizes
- End-to-End Learning - Train entire pipeline jointly
Key Tasks
Image Classification
Assign label to entire image.
Architectures:
- ResNet: Residual connections for deep networks
- EfficientNet: Optimal scaling
- Vision Transformers: Attention-based
Approach:
- Use pre-trained model
- Fine-tune on target dataset
- Data augmentation
Object Detection
Locate and classify objects in images.
Methods:
- YOLO: Single-stage, real-time
- Faster R-CNN: Two-stage, accurate
- RetinaNet: Focal loss for imbalance
Components:
- Backbone: Feature extraction
- Neck: Feature fusion
- Head: Detection and classification
Image Segmentation
Classify each pixel.
Types:
- Semantic: Class per pixel
- Instance: Separate object instances
- Panoptic: Semantic + instance
Architectures:
- U-Net: Encoder-decoder with skip connections
- Mask R-CNN: Instance segmentation
- DeepLab: Atrous convolutions
Using the Reference Files
/references/classification.md — CNN architectures, transfer learning, fine-tuning strategies, data augmentation, and evaluation metrics.
/references/object-detection.md — YOLO, Faster R-CNN, anchor boxes, non-maximum suppression, evaluation (mAP), and real-time detection.
/references/segmentation.md — U-Net, Mask R-CNN, semantic vs instance segmentation, loss functions, and post-processing.
Best Practices
- Use pre-trained models (ImageNet weights)
- Apply data augmentation (crop, flip, color jitter)
- Normalize inputs (ImageNet mean/std)
- Use appropriate image size for model
- Monitor for overfitting
- Evaluate on diverse test set
- Consider inference speed requirements
- Handle class imbalance
Common Pitfalls to Avoid
- Not using pre-trained weights
- Insufficient data augmentation
- Wrong input normalization
- Inappropriate image resolution
- Ignoring class imbalance
- Not considering inference speed
- Overfitting to small datasets
- Poor anchor box design (object detection)