**CornerNet: Detecting Objects as Paired Keypoints** **Authors:** Hei Law, Jia Deng **Abstract:** CornerNet is a novel approach to object detection that detects bounding boxes as pairs of keypoints—specifically, the top-left and bottom-right corners. By using a single convolutional neural network (CNN), CornerNet eliminates the need for anchor boxes, which are commonly used in single-stage detectors. The paper introduces corner pooling, a new type of pooling layer that helps the network better localize corners. Experiments show that CornerNet achieves a 42.2% average precision (AP) on the Microsoft COCO dataset, outperforming all existing one-stage detectors. **Keywords:** Object Detection **Introduction:** The paper discusses the limitations of anchor boxes in object detection, such as the large number of anchor boxes required and the hyperparameters involved. CornerNet proposes a new approach that detects objects as pairs of keypoints, simplifying the network output and eliminating the need for anchor boxes. The method is inspired by associative embedding, which is used for multi-person pose estimation. CornerNet uses an hourglass network as its backbone and includes corner pooling to enhance corner localization. The paper also introduces a modified focal loss to improve training efficiency. **Related Works:** The paper reviews two-stage and one-stage object detectors, highlighting the advantages and challenges of each approach. It discusses the limitations of anchor boxes in one-stage detectors and introduces CornerNet as a solution. **CornerNet:** CornerNet detects objects as pairs of keypoints using a single CNN. It predicts heatmaps for top-left and bottom-right corners, embeddings for each detected corner, and offsets to adjust corner locations. The network is trained to predict similar embeddings for corners from the same object. Corner pooling is used to better localize corners by encoding explicit prior knowledge. **Experiments:** The paper evaluates CornerNet on the MS COCO dataset, demonstrating its effectiveness through ablation studies and comparisons with other detectors. Corner pooling and the hourglass network are found to be crucial for performance. The main bottleneck in CornerNet is detecting corners, as shown by error analysis. **Conclusion:** CornerNet is a novel approach to object detection that detects bounding boxes as pairs of keypoints, eliminating the need for anchor boxes. It achieves competitive results on the MS COCO dataset, outperforming existing one-stage detectors.**CornerNet: Detecting Objects as Paired Keypoints** **Authors:** Hei Law, Jia Deng **Abstract:** CornerNet is a novel approach to object detection that detects bounding boxes as pairs of keypoints—specifically, the top-left and bottom-right corners. By using a single convolutional neural network (CNN), CornerNet eliminates the need for anchor boxes, which are commonly used in single-stage detectors. The paper introduces corner pooling, a new type of pooling layer that helps the network better localize corners. Experiments show that CornerNet achieves a 42.2% average precision (AP) on the Microsoft COCO dataset, outperforming all existing one-stage detectors. **Keywords:** Object Detection **Introduction:** The paper discusses the limitations of anchor boxes in object detection, such as the large number of anchor boxes required and the hyperparameters involved. CornerNet proposes a new approach that detects objects as pairs of keypoints, simplifying the network output and eliminating the need for anchor boxes. The method is inspired by associative embedding, which is used for multi-person pose estimation. CornerNet uses an hourglass network as its backbone and includes corner pooling to enhance corner localization. The paper also introduces a modified focal loss to improve training efficiency. **Related Works:** The paper reviews two-stage and one-stage object detectors, highlighting the advantages and challenges of each approach. It discusses the limitations of anchor boxes in one-stage detectors and introduces CornerNet as a solution. **CornerNet:** CornerNet detects objects as pairs of keypoints using a single CNN. It predicts heatmaps for top-left and bottom-right corners, embeddings for each detected corner, and offsets to adjust corner locations. The network is trained to predict similar embeddings for corners from the same object. Corner pooling is used to better localize corners by encoding explicit prior knowledge. **Experiments:** The paper evaluates CornerNet on the MS COCO dataset, demonstrating its effectiveness through ablation studies and comparisons with other detectors. Corner pooling and the hourglass network are found to be crucial for performance. The main bottleneck in CornerNet is detecting corners, as shown by error analysis. **Conclusion:** CornerNet is a novel approach to object detection that detects bounding boxes as pairs of keypoints, eliminating the need for anchor boxes. It achieves competitive results on the MS COCO dataset, outperforming existing one-stage detectors.