CCFF boosts autonomous driving detection with 14.1% gain on small objects
New attention-based framework achieves 0.973 consistency score while running in real-time.
Object detection in autonomous driving remains challenging due to rare classes, small-scale objects, and complex co-occurrence patterns. A new paper from researchers Binay Kumar Singh and Niels Da Vitoria Lobo introduces Context-Centric Feature Fusion (CCFF), a framework that integrates two attention-based modules to handle these issues. The Local Context Fusion Module (LCFM) uses RoI-to-RoI self-attention to resolve spatial interactions for small and partially obscured objects. The Global Context Attention Module (GCAM) pools top-K RoI features into a global context token, capturing object co-occurrence without pixel-level global pooling overhead. This fusion produces contextualized embeddings that significantly enhance classification and detection of co-occurring objects.
Evaluated on Cityscapes and BDD100K, CCFF achieves Category-level Consistency Strategy (CCS) scores of 0.973 and 0.969 respectively, demonstrating strong relational consistency. It yields a substantial 14.1% improvement in small object detection (AP_S) and successfully recovers rare classes like 'Train' that are typically lost in large distributions. The framework processes images in real-time with only a 0.2 FPS overhead, making it practical for deployment in autonomous vehicles. Code is available on GitHub. The paper will be presented at the CVPR 2026 Precognition Workshop.
- CCFF uses LCFM (local RoI-to-RoI attention) for small/occluded objects and GCAM (global context token) for co-occurrence priors.
- Achieves 0.973 CCS on Cityscapes and 0.969 on BDD100K, with 14.1% AP improvement on small objects.
- Runs in real-time with only 0.2 FPS overhead, recovering rare classes like 'Train' frequently missed by standard detectors.
Why It Matters
Enables autonomous vehicles to detect small, rare, and co-occurring objects reliably without sacrificing real-time performance.