Accelerating Non-Maximum Suppression: A Graph Theory Perspective

Authors: King-Siong Si, Lu Sun, Weizhan Zhang, Tieliang Gong, Jiahao Wang, Jiang Liu, Hao Sun

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Taking the YOLOv8-N model on MS COCO 2017 as the benchmark setup, our method QSI-NMS provides 6.2 speed of original NMS on the benchmark, with a 0.1% decrease in m AP.
Researcher Affiliation Collaboration 1School of Computer Science and Technology, MOEKLINNS Lab, Xi an Jiaotong University 2School of Computer Science and Technology, BDKE Lab, Xi an Jiaotong University 3Institute of Artificial Intelligence (Tele AI), China Telecom
Pseudocode Yes The pseudo-code for QSI-NMS can be found in the Appendix. The pseudo-code for BOE-NMS is described in Algorithm 3 which can be found in the Appendix. Appendix E.1 Pseudo-Code for QSI-NMS (Algorithm 1), E.2 Pseudo-Code for e QSI-NMS (Algorithm 2), E.3 Pseudo-Code for BOE-NMS (Algorithm 3).
Open Source Code Yes The code for NMS-Bench is available on Git Hub .
Open Datasets Yes We conduct tests on MS COCO 2017 [24] and Open Images V7 [23] using YOLOv5 [20], YOLOv8 [21], and Faster R-CNN [12] as validation models.
Dataset Splits Yes Figure 2: Statistical characteristics of graph G on MS COCO 2017 validation.
Hardware Specification Yes Our experimental environment is shown as the Table 7. Table 7: Experimental Environment Component Specification Model Intel Xeon Gold 6226 Total Cores 12 Total Threads 24 Max Turbo Frequency 3.70 GHz Model NVIDIA RTX 4090 1 VRAM 24 GB GDDR6X
Software Dependencies No The paper mentions implementing methods as C++ operators under the torchvision library, but does not provide specific version numbers for these software components.
Experiment Setup Yes For the hyperparameter settings, we set the NMS threshold Nt to 0.7 in our experiments. We set bench size as 20.