Overturning the Counting Cornerstone: Exploring Fine-Grained Adaptive Losses to Subvert the Conventional Density Estimation

Previous works of crowd counting prepossess the given label and convert it into a density map or count map used for learning. However, we revealed that density maps tend to have severe errors due to faulty occlusions, head size variation, and head shape variation. Directly learning the density map will often result in fatal over-fitting. On the other hand, Count-map did not fully utilize the detailed information of the image. These unsatisfactory preprocessing lead to the performance bottleneck despite recent advances in network architecture. To solve these problems, in this paper, we discovered that the distribution of errors throughout the density map is not uniform. Moreover, it is correlated with the distance to the nearest annotation point. Inspired by this finding, we introduce Fine-Grained Adaptive Losses to learn the density map differently in different regions of the density map. While our method is simple, it dictates that we should endeavor to obtain more supervision from the density map. Our effort subverts the traditional use of density maps and opens up a new vision for future counting research. Extensive experiments demonstrate that our approach significantly outperforms standard methods in crowd-counting datasets.