FusionCount: Efficient Crowd Counting via Multiscale Feature Fusion
Why build an extra multiscale module if the encoder has already produced features with many receptive-field sizes?
01 / Problem
What is this work trying to solve?
Perspective makes people occupy very different image scales. Earlier counters address this by running extra branches or spatial-pyramid modules after the encoder, but those modules add computation, cover only a few predefined receptive fields, and ignore the multiscale information already present in shallow and intermediate encoder features.
02 / Idea
What is the central technical intuition?
FusionCount collects VGG-16 features after the first pooling stage, spanning receptive fields from 6 to 192 pixels. Within each resolution group, contrast features generate spatial weights for adaptive averaging; the four fused maps are then decoded coarse-to-fine. A parallel dilated-convolution and pointwise-convolution block reduces channels while retaining salient context.
03 / Evidence
What does the current evidence show?
On ShanghaiTech B, FusionCount reports 6.9 MAE and 11.8 RMSE, the best MAE and tied-best RMSE among the compared VGG-based models. On the harder ShanghaiTech A it is competitive rather than best at 62.2 MAE / 101.2 RMSE. Its reported cost is 815.0G multiply-adds at 1920 x 1080, below CSRNet, CAN, BL, and DM-Count in the paper's comparison.
04 / Limits
Where are the limits of this work?
The evaluation covers only ShanghaiTech A and B, and efficiency is measured as multiply-adds rather than wall-clock latency, memory use, or energy. The decoder's second-stage fusion also uses simple addition; the paper leaves context-aware decoding fusion to future work.