Summary
This paper presents a new object detector YOLOv6 for real-time object detection. The technical novelties consist of a new network design, an anchor-based test-free auxiliary branch for training, and an adaptive self-distillation for transferring knowledge from a teacher to a student. Experimental results show the state-of-the-art performance of the proposed YOLOv6 among various YOLO series.
Weaknesses
1. While the proposed method achieves state-of-the-art performance on COCO val2017, the paper does not show the results on the COCO test-dev set. Since this detector is made for pursuing SoTA performance in object detection, it should provide the evaluation on the COCO test-dev set.
2. The paper introduces several strategies for improving detection performance, some of which are not cost-free in training, e.g., anchor-aided training and self-distillation. However, the paper does not provide the results for training time cost.
3. The comparison is unfair. The self-distillation requires a pre-trained teacher model, which makes the YOLOv6 need double the training time. In contrast, the previous works like YOLOv5, YOLOv8, and DETR series do not need that. I noticed that the paper conducts an experiment to check the performance of double training epochs, showing the superiority of self-distillation. However, in the benchmark test, all the detectors train for 300 epochs, while YOLOv6 actually trains for 600 epochs (300 for teacher and 300 for student). This is a serious problem which makes the comparison unfair.
4. It is shown in Table 17 that SIoU or CIoU loss works better than GIoU. Then why did you choose a sub-optimal one GIoU?
5. The label assignment algorithm TAL, VFL loss, and GIoU loss mentioned in Sec. 4.3.1 lack references in the main body of the paper, though they appear in the Appendix.
6. The cosine weight decay sounds like a regularization while it is not in the proposed self-distillation method. I suggest the author choose a better name for it.
7. There are some inappropriate statements and ambiguities that may be misleading to readers. 1) In Sec. 3.3, you mentioned: "our large models (i.e., YOLOv6-M/L) adopt DFL [1] **as regression loss** for the convenience of **performing self-distillation** on localization". As a matter of fact, DFL is not a box regression loss. It should be the IoU-based losses. DFL loss is a weighted cross-entropy loss, which is proposed as additional supervision to make the convergence faster and more stable. 2) DFL is not originally proposed for distillation. It should be the localization distillation [2]. 3) On the last paragraph of page 5, "Notably, the introduction of DFL (Li et al., 2020) requires extra parameters for the regression branch, ..." DFL (distribution focal loss) is just a loss function. It should be the general distribution representation of the bounding box. 4) What is the experimental setting of double epochs in Table 10?
[1] Li X, Wang W, Wu L, et al. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection[J]. Advances in Neural Information Processing Systems, 2020, 33: 21002-21012.
[2] Zheng Z, Ye R, Hou Q, et al. Localization distillation for object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023.
Confidence
4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.