DensFiLM: Density-Conditioned Video Saliency for Crowd Scenes

Video saliency models typically treat all crowd scenes with a uniform fixation strategy. However, human attention shifts significantly depending on crowd density: in sparse scenes, observers tend to track individual trajectories, whereas in dense or congested scenes, attention shifts toward collective motion, group boundaries, and scene-level landmarks. This paper investigates whether conditioning a video saliency model on crowd density can improve performance by allowing the network to adapt its feature processing to these distinct visual regimes.

The authors introduce DensFiLM, a density-conditioned model built upon a Video Swin Transformer backbone. The core innovation is the insertion of a Feature-wise Linear Modulation (FiLM) layer at the encoder's bottleneck. This module uses a learned density embedding to apply channel-wise scale and shift parameters to the feature map. This approach is highly efficient, adding only approximately 100,000 parameters. The model can operate using either ground-truth density labels or its own auxiliary density predictions, making it practical for real-world inference without manual annotation.

DensFiLM achieves state-of-the-art results on the CrowdFix benchmark, significantly outperforming the previous baseline, ACLNet. Notably, the model's performance when using its own predicted density labels is nearly identical to its performance when using oracle (ground-truth) labels. The authors also conducted diagnostic tests showing that density conditioning provides a substantial gain in Normalized Scanpath Saliency (NSS) compared to unconditioned backbones. Interestingly, more complex architectural extensions—such as incorporating explicit RAFT optical flow or social force priors—failed to improve performance, suggesting that for the available training data, a targeted inductive bias is more effective than simply increasing model capacity.