NOH-NMS: Improving Pedestrian Detection by Nearby Objects Hallucination
This addresses pedestrian detection in dense scenarios, offering a specific improvement over existing NMS methods.
The paper tackled the problem of greedy-NMS causing low recall or high false positives in pedestrian detection, especially in dense crowds, by proposing NOH-NMS, which dynamically eases suppression based on nearby object likelihood, resulting in improvements of 3.9% AP, 5.1% Recall, and 0.8% MR^{-2} on CrowdHuman.
Greedy-NMS inherently raises a dilemma, where a lower NMS threshold will potentially lead to a lower recall rate and a higher threshold introduces more false positives. This problem is more severe in pedestrian detection because the instance density varies more intensively. However, previous works on NMS don't consider or vaguely consider the factor of the existent of nearby pedestrians. Thus, we propose Nearby Objects Hallucinator (NOH), which pinpoints the objects nearby each proposal with a Gaussian distribution, together with NOH-NMS, which dynamically eases the suppression for the space that might contain other objects with a high likelihood. Compared to Greedy-NMS, our method, as the state-of-the-art, improves by $3.9\%$ AP, $5.1\%$ Recall, and $0.8\%$ $\text{MR}^{-2}$ on CrowdHuman to $89.0\%$ AP and $92.9\%$ Recall, and $43.9\%$ $\text{MR}^{-2}$ respectively.