Cross-Prior Reliability Fusion for Traffic Salient Object Boundary Detection in Autonomous Driving

Technical Summary: Cross-Prior Reliability Fusion for Traffic Salient Object Boundary Detection

Problem Statement
Traffic Salient Object Detection (TSOD) in autonomous driving requires a dual capability: identifying complete object regions (e.g., vehicles, pedestrians) and achieving precise contour localization. Existing approaches face a fundamental spatial reliability conflict. Semantic region prediction models (e.g., DeepLabV3) excel at capturing complete object interiors and suppressing irrelevant background structures but often produce over-smoothed, ambiguous boundaries. Conversely, high-resolution boundary models (e.g., ResNet50-FPN) preserve fine-grained structural details but are prone to false positives triggered by non-target visual structures such as lane markings, reflections, and texture edges.

Traditional end-to-end fusion strategies often rely on joint optimization of feature representations, which can be data-hungry and may bias models toward dense region information at the expense of sparse boundary pixels. Furthermore, fixed fusion weights fail to adapt to the varying reliability of predictions across different spatial locations within complex road environments.

Methodology: Cross-Prior Reliability Fusion (CPRF)
The authors propose Cross-Prior Reliability Fusion (CPRF), a parameter-efficient framework that reformulates boundary detection as a prediction-level reliability calibration problem rather than a network design problem. The core philosophy is to leverage complementary priors from frozen models and learn lightweight spatial corrections.

1. Frozen Complementary Prior Extraction:
   CPRF utilizes two pre-trained, frozen extractors:

   • A Semantic Prior Extractor (DeepLabV3-ResNet50) to generate region logits ($M_s$).
   • A Boundary Prior Extractor (ResNet50-FPN) to generate region ($M_c$) and boundary ($E_c$) logits.
     During training, the parameters of these base models ($\theta_S, \theta_C$) remain frozen; only the calibration modules are updated.

2. Reliability Evidence Construction:
   To determine which prediction source is more reliable at specific spatial locations, CPRF constructs three types of evidence:

   • Semantic Gradient ($G_s$): Derived from max/min pooling of semantic probabilities to identify transition regions.
   • Normalized Uncertainty Proxy ($U_s$): Calculated as $4P_s(1-P_s)$ to identify regions where the semantic prediction is ambiguous (near 0.5 probability).
   • Cross-Prior Consistency ($A$): An exponential map of the absolute difference between semantic and boundary region predictions, indicating agreement or disagreement between the two priors.

3. Spatial Reliability Calibration:

   • Fixed Prior Construction: An initial boundary prior is created via equal-weight fusion of the semantic and boundary logits to establish a stable baseline.
   • Boundary Calibration: A lightweight calibrator ($C_e$) takes the fixed prior and the reliability evidence as input to predict bounded spatial residuals. The output is constrained by a tanh function to prevent excessive deviation from the fixed fusion result.
   • Mask Calibration: An auxiliary pathway ($C_m$) ensures region completeness is preserved during boundary refinement, using consistency constraints and distance-based optimization.

4. Training Objective:
   The framework optimizes a multi-term loss function comprising:

   • Mask and Edge losses (BCE and Dice) for region and boundary accuracy.
   • Consistency Loss ($L_{cons}$) to align the spatial gradients of the calibrated mask and boundary.
   • Distance Loss ($L_{dist}$) to penalize boundary responses far from ground-truth contours.

Key Contributions

1. Problem Reformulation: The work shifts the focus from designing new end-to-end architectures to treating TSOD as a spatial reliability calibration problem between complementary semantic and boundary priors.
2. Parameter-Efficient Framework: CPRF introduces a calibration mechanism that optimizes only 7,754 parameters (0.0119% of the total 65.41M parameters) while keeping the underlying visual prior extractors frozen. This preserves the inductive biases of pre-trained models and mitigates catastrophic forgetting.
3. Coupled Calibration Mechanism: The design includes a coupled mask calibration pathway that jointly optimizes object region completeness and contour localization, preventing the trade-off where boundary sharpening degrades region integrity.
4. Spatially-Aware Evidence: Unlike global confidence calibration, CPRF learns image-conditioned local calibration strategies using semantic gradients, uncertainty proxies, and cross-prior consistency to adaptively adjust prediction reliability.

Experimental Results
Experiments were conducted on the TSOD10K dataset (9,897 training, 1,100 validation, 2,756 test images).

• Performance: CPRF achieved an Optimal Dataset Scale (ODS) of 0.5437, Optimal Image Scale (OIS) of 0.5804, and Average Precision (AP) of 0.4988.
• Improvement: These results represent improvements over the equal-weight fusion baseline of +0.0085 (ODS), +0.0066 (OIS), and +0.0029 (AP).
• Stability: Performance remained consistent across three random seeds (48, 49, 50), with standard deviations as low as 0.0004 for OIS.
• Data Efficiency: The model achieved significant gains even when trained on only 25% of the calibration data, demonstrating that improvements stem from the calibration mechanism rather than large-scale retraining.
• Complexity: While the total parameter count is 65.40M (due to frozen backbones), the trainable ratio is extremely low (0.0119%). Inference speed was measured at 40.15 FPS on an NVIDIA RTX 3080 Ti.

Significance and Claims
The paper claims that the primary challenge in traffic salient object boundary detection is not merely generating stronger edge responses, but determining the relative reliability of different prediction sources at different spatial locations.

The authors emphasize that CPRF does not aim to be a lightweight deployment model for embedded systems, as it still requires the computational cost of running two frozen prior extractors during inference. Instead, its significance lies in parameter-efficient adaptation: it demonstrates that high-quality boundary calibration can be achieved by learning a small number of spatial residuals on top of existing, frozen visual priors. This approach allows for task adaptation in data-limited scenarios without the risk of representation degradation associated with full model retraining. The work suggests that spatial reliability calibration is a viable strategy for improving autonomous driving perception without updating the underlying visual feature extractors.