EIVE: End-to-End Instance-Specific Visual Explanations for Detection Transformers
Researchers have developed EIVE, a novel framework for generating instance-specific visual explanations for object detection models like DETR. Unlike existing post-hoc methods that require extra computation, EIVE directly produces saliency maps during the model's forward pass by reformulating its cross-attention mechanism. This approach enhances computational efficiency and can be applied to various DETR-like architectures, with experiments showing competitive or superior performance in explanation quality and detection accuracy. AI
IMPACT Enhances interpretability of object detection models, potentially improving debugging and trust in AI systems.