We sincerely appreciate the time and effort you have invested in reviewing our rebuttal, and we are pleased that our response has addressed your main concern.
We are also more than happy to provide a comparison of the mAP performance comparison between our methods and PSRR-MaxpoolNMS. We apologize for being unable to fully present this in the rebuttal, as we had reached the character limit. We tested anchor-based models (Faster R-CNN, YOLOv5) where PSRR-MaxpoolNMS is applicable, using the MS COCO dataset. Please see the results in the tables below.
Table 1: Faster R-CNN R50-FPN (average #bounding boxes: 251)
||Original NMS|PSRR-MaxpoolNMS|BOE-NMS|QSI-NMS|eQSI-NMS|
| ---------------- | ------------ | ---- | ------- | ------- | -------- |
| mAP(%) |39.8| 37.5 | 39.8 | 39.5 | 39.3 |
| latency($\mu s$) |53.0| 89.0 | 43.1 | 34.3 | 24.6 |
Table 2: Faster R-CNN R101-FPN (average #bounding boxes: 236)
||Original NMS|PSRR-MaxpoolNMS|BOE-NMS|QSI-NMS|eQSI-NMS|
| ---------------- | ------------ | ---- | ------- | ------- | -------- |
| mAP(%)|41.8|39.5|41.8|41.5|41.4|
| latency($\mu s$)|45.5|86.4|38.6|31.9|23.0|
Table 3: Faster R-CNN X101-FPN (average #bounding boxes: 214)
|| Original NMS | PSRR-MaxpoolNMS | BOE-NMS | QSI-NMS | eQSI-NMS |
| ---------------- | ------------ | --------------- | ------- | ------- | -------- |
| mAP(%)| 43.0| 40.5| 43.0| 42.7| 42.5|
| latency($\mu s$) | 37.0| 86.9| 33.4| 28.6| 20.6|
Table 4: YOLOv5-N (average #bounding boxes: 2898)
|| Original NMS | PSRR-MaxpoolNMS | BOE-NMS | QSI-NMS | eQSI-NMS |
| ---------------- | ------------ | --------------- | ------- | ------- | -------- |
| mAP(%)| 27.8| 26.5| 27.8| 27.5|27.4|
| latency($\mu s$)|8568.5| 599.6| 906.2|628.0|325.6|
Table 5: YOLOv5-S (average #bounding boxes: 1974)
| | Original NMS | PSRR-MaxpoolNMS | BOE-NMS | QSI-NMS | eQSI-NMS |
| ---------------- | ------------ | --------------- | ------- | ------- | -------- |
| mAP(%) | 37.2| 35.6 | 37.2 | 36.9 | 36.6 |
| latency($\mu s$) | 3858.2 | 409.4 | 547.7 | 408.2 | 217.5 |
Table 6: YOLOv5-M (average #bounding boxes: 1810)
| | Original NMS | PSRR-MaxpoolNMS | BOE-NMS | QSI-NMS | eQSI-NMS |
| ---------------- | ------------ | --------------- | ------- | ------- | -------- |
| mAP(%)| 45.1 | 43.1 | 45.1 | 44.9 | 44.5 |
| latency($\mu s$) | 2918.1| 380.5 | 424.7 | 371.3 | 197.4|
As you can see, eQSI-NMS achieves the lowest latency while maintaining a favorable trade-off with mAP. However, PSRR-MaxpoolNMS experiences a $1–2$% mAP accuracy loss in both the Faster R-CNN and YOLOv5 models. This is likely due to the following reasons:
1. Some hyperparameters need adjustment when using the MS COCO dataset, as MS COCO is more complex compared to PASCAL VOC, with a richer variety of categories and broader scenes. Therefore, the number of scales and ratios may need to be increased to suit the MS COCO dataset. In contrast, our proposed methods do not require additional parameters beyond those used in Original NMS, making them more applicable to general cases.
2. The PSRR-MaxpoolNMS paper does not mention how the input image size and the settings of $W$ and $H$ in PSRR-MaxpoolNMS are determined. In our implementation, we set this to $640 \times 640$ to accommodate all images in the MS COCO dataset. This might affect accuracy, though we believe the impact is minimal. This also indicates that our method offers better generalizability.
In the case of Faster R-CNN, the latency performance of PSRR-MaxpoolNMS is not competitive. This is because PSRR-MaxpoolNMS requires 8 maxpooling operations, which, although not affecting the algorithm's complexity, introduces a large constant factor that hampers efficiency when the number of bounding boxes is small (e.g., the average number of bounding boxes in the three Faster R-CNN models is less than 300). However, it performs well when the number of bounding boxes is large (e.g., YOLOv5-S has an average of 2898 bounding boxes). This demonstrates that the speedup of PSRR-MaxpoolNMS is highly dependent on the degree of parallelism, whereas our method directly reduces computational overhead (see Figure 6 in Appendix D.3), making it hardware-agnostic and suitable for resource-constrained edge devices.
In summary, the limitations of PSRR-MaxpoolNMS in improving efficiency while balancing accuracy are significant. In contrast, our method, which uses the same inputs and parameters as Original NMS, is a plug-and-play algorithm that can directly replace Original NMS. To demonstrate this, we implemented our method in the torchvision library and compared it with the highly parallel CUDA NMS [1], where our method exhibited significant superiority (see Table 7 in Appendix D.3).
Thanks again for the time you invested in writing your comments. We hope this response can thoroughly address your concern.
---
[1] TorchVision maintainers and contributors. TorchVision: PyTorch’s Computer Vision library, November 2016.