Abstract
Multi-class part parsing in "in the wild" scenarios presents significant challenges due to part-level boundary blurring and inter-class semantic ambiguity. This paper introduces BSANet, a framework that integrates lightweight Boundary Awareness and Semantic Awareness modules to address these issues. By leveraging spatial boundary attention and semantic channel selection, BSANet achieves state-of-the-art performance on the Pascal-Part dataset while maintaining high inference speeds, processing images roughly 30x faster than previous leading methods.
Introduction
Traditional semantic part parsing often assumes a single, well-localized object category. However, real-world images typically contain multiple objects of diverse classes. This "multi-class" setting introduces two primary hurdles:
- Part-level Ambiguity: Downsampling in deep networks often blurs the precise boundaries between small, adjacent parts (e.g., a bird's beak or a human's hand).
- Class-level Ambiguity: Visually similar parts across different categories (e.g., a bus front vs. a car front) lead to inter-class confusion.
The BSANet Framework
BSANet, built upon a DeepLabV3 backbone, introduces two plug-and-play modules to resolve these ambiguities:
Boundary Awareness Module (BAM)
The BAM focuses on spatial localization. It utilizes an auxiliary edge encoder to predict class-agnostic part boundaries. This prediction serves as a spatial attention mask, forcing the network to prioritize transitions between parts. These refined features are fused into a cascade pyramid to preserve fine-grained details.
Semantic Awareness Module (SAM)
The SAM addresses "what" is being parsed. A semantic branch predicts the overall object category at each pixel. This information acts as a prior, performing channel-wise attention to "switch on" feature channels relevant to the detected class and "switch off" irrelevant ones, thereby preventing misclassification between similar-looking parts of different objects.
Experimental Results
The model was evaluated on the Pascal-Part and Pascal-Person-Part datasets.
Performance and Efficiency
BSANet achieved 58.2% mIoU on the full Pascal-Part benchmark, a significant jump from the 54.4% baseline. On the Pascal-Person-Part subset, it reached a record 68.43% mIoU, outperforming specialized models that utilize additional data like human pose skeletons.
Crucially, BSANet is highly efficient. It processes a 512x512 image in under 200ms on an NVIDIA 1080Ti, making it approximately 30x faster than previous state-of-the-art models like HAZN.
Conclusion
By treating boundaries as spatial filters and object categories as semantic priors, BSANet effectively tackles the complexities of multi-class part parsing. Its lightweight design proves that high-level structural modeling is not always necessary for state-of-the-art performance, offering a practical and efficient solution for real-world computer vision tasks.
Summary updated