Week 1 · Day 3 · Wednesday 14 October

Object Detection and Segmentation

Finding and outlining objects: IoU, region proposals, the R-CNN and YOLO families, then semantic, instance and panoptic segmentation with FCNs and U-Net.

Topic Description & Objectives Resource
Object Detection IoU, detecting one and then many objects, region proposals, the R-CNN family, single-shot detectors and YOLO, and an industry use case Open
Image Segmentation Semantic, instance and panoptic segmentation, upsampling and transposed convolutions, FCNs and U-Net, keypoints and pose, and object tracking Open
Lab 1: Object Detection with Faster R-CNN Fine-tune a pretrained Faster R-CNN on Pascal VOC, with a custom Dataset class and collate function, and visualise predictions against ground truth Notebook Colab
Lab 2: Object Detection with YOLOv8 Detect wind turbines with Ultralytics YOLOv8: the YOLO label format, training, validation and inference Notebook Colab
Lab 3: Segmentation with a Custom U-Net Build a U-Net with residual connections and train it to segment pets in the Oxford-IIIT Pet dataset Notebook Colab
Optional: U-Net with a Pretrained Encoder Segment flooded areas with segmentation_models_pytorch and a pretrained encoder Notebook Colab
Optional: Object Detection with RT-DETR The wind-turbine detector again, with the transformer-based RT-DETR Notebook Colab
Optional: Segment Anything 3 Run Meta's Segment Anything 3 through Hugging Face Transformers to segment objects in images Notebook Colab