📄️ Advanced computer vision
From an image to a pipeline that counts vehicles and pedestrians at a crossroads: two-stage and one-stage detection, segmentation, multi-object tracking, COCO metrics and a final project.
📄️ 1. Three families of tasks
Module 1 of the advanced computer vision premium course: distinguish classification, detection and segmentation, pick the right annotation format and estimate its true cost.
📄️ 2. Two-stage detection (R-CNN)
Module 2 of the advanced computer vision premium course: region proposals, RoI pooling and RoI Align, and the accuracy-versus-speed trade-off of the R-CNN family.
📄️ 3. One-stage detection (YOLO, SSD)
Module 3 of the advanced computer vision premium course: dense predictions on a grid, the YOLO family from v3 to v8, SSD and RetinaNet's focal loss, and the real-time constraint.
📄️ 4. Anchors, NMS and thresholds
Module 4 of the advanced computer vision premium course: how anchors bias what a detector sees, how NMS collapses duplicates, and how to set confidence and IoU thresholds without wrecking precision or recall.
📄️ 5. Detection metrics (IoU, mAP)
Module 5 of the advanced computer vision premium course: compute IoU by hand, build a precision-recall curve, and read a COCO report from mAP@0.5 to mAP@[0.5:0.95].
📄️ 6. Semantic segmentation
Module 6 of the advanced computer vision premium course: encoder-decoder with skip connections, atrous convolutions, Dice and per-class IoU, and how to fight pixel class imbalance.
📄️ 7. Instance segmentation
Module 7 of the advanced computer vision premium course: adding a mask head to Faster R-CNN, differences with semantic segmentation, panoptic as a bridge, and Segment Anything as an annotation booster.
📄️ 8. Multi-object tracking
Module 8 of the advanced computer vision premium course: associate detections to tracks with the Hungarian algorithm, predict motion with a Kalman filter, and count crossings with SORT or ByteTrack.
📄️ 9. Annotation and augmentation
Module 9 of the advanced computer vision premium course: annotation tools and inter-annotator agreement, box-aware augmentation with Albumentations, and near-duplicate leakage between train and test.
📄️ 10. Project: custom detection
Module 10 of the advanced computer vision premium course: the full pipeline from a custom crossroads dataset to a fine-tuned YOLOv8, evaluated with COCO metrics and analysed by class and object size.
📄️ Recap and exam
Complete recap of the advanced computer vision premium course: detection, segmentation, tracking, dataset quality and the crossroads project, then the 40-question exam.