Core operations
- Convolution from Scratch: im2col lowering, stride, padding, dilation
- Pooling: max and average pooling, backward pass
- Bilinear Resize: coordinate mapping and four-tap interpolation
- Image Padding and Flips: boundary modes, mirror index arithmetic
- Batched Image Preprocessing: normalize, channel order, letterbox
- Edge Detection: gradients, Sobel, Canny, Laplacian of Gaussian
Representation and architectures
- Vision Transformers: patch tokens through a transformer encoder
- Patch Embedding: image to tokens, positions, CLS
- Feature Pyramids: top-down multi-scale backbone features (FPN)
- CLIP (Contrastive Language-Image Pretraining): contrastive image-text pretraining, zero-shot transfer
- Perceptual Distance: LPIPS-style distance in deep features
Detection and segmentation
- IoU and Box Metrics: IoU, GIoU, mAP at a glance
- Non-Maximum Suppression: greedy NMS and soft-NMS decay
- Semantic Segmentation Metrics: pixel accuracy, mIoU, Dice score
- Open-Vocabulary Detection: detecting classes named at test time
Generative evaluation
- FID: Frechet distance between Inception feature Gaussians
- Inception Score: classifier confidence times label distribution diversity
- FVD: FID over I3D spatiotemporal features
- PSNR and SSIM: log MSE and windowed structural similarity
Video
- Video Tokenization: spatiotemporal VQ-VAE and compression tradeoffs
- Spatiotemporal Attention: full versus factorized space-time attention cost
- Frame Sampling and Preprocessing: fps, clip sampling, temporal aggregation
Augmentation and robustness
- Image Augmentation: crops, jitter, cutout, mixup, cutmix
- PCA Color Augmentation: color jitter along RGB principal axes
Anomaly and outliers
- Isolation Forest: random splits isolate anomalies in few cuts
- Local Outlier Factor: local density ratio against k neighbors
- One-Class SVM: max-margin boundary separating data from origin
3D
- NeRF: volumetric radiance field from posed images
4 items under this folder.