# DETR - Transformer architecture for object detection - end-to-end object detection model - consider object detection as a direct set prediction problem --- - **No NMS (non-maximum supression)** - two-stage detectors predict boxes w.r.t proposals - eliminate redundant overlapping bounding boxes. - **No anchors generation** - single-stage detectors predict boxes w.r.t anchors - To provide a set of predefined boxes that the model adjusts to match the ground truth boxes during training --- - NMS: bipartite matching loss, no object - anchor: object queries  --- # DETR model - object detection set prediction loss - DETR architecture  --- - set prediction loss  --- - $l_1$ loss will have different scales for small and large boxes even if their relative errors are similar - so add generalized IOU loss  --- - DETR perform well on large objects but still have limitations on small object at that time.  --- - encoder self-attention  --- - decoder self-attention  --- - green : small boxes - red: large horizontal boxes - blue: large vertical boxes  --- - DETR for panoptic segmentation  --- - FPN (Feature Pyramid Network) - commonly used with CNN? - MLA (multi-level features aggregation) - commonly used with transformer? --- - architecture of segformer  --- ---