[Task] object-detection EP coverage investigation

Overview

Object detection models predict bounding boxes and class labels for objects in images. This task covers table structure detection (table-transformer), real-time detection (RT-DETR), and YOLO-style transformers (YOLOS), plus fashion-domain detection. All 10 models pass on at least one EP.

This issue focuses on the 5 models with incomplete EP coverage — primarily a VitisAI failure pattern (4 models), plus one model failing on both QNN and VitisAI. Root cause investigation needed across DETR and RT-DETR architectures.

Agent Scenarios

• Document intelligence agent: extract table structure from scanned documents or PDFs using table-transformer models, enabling structured data extraction from unstructured layouts
• Retail / inventory agent: detect product locations in warehouse or shelf images (yolos-fashionpedia, RT-DETR) for automated inventory tracking
• Visual copilot agent: ground natural language object references to bounding boxes in a scene, enabling spatial reasoning for multimodal agents

EP Coverage Status

Reference: microsoft/table-transformer-structure-recognition, hustvl/yolos-small, facebook/detr-resnet-50, valentinafeve/yolos-fashionpedia, TahaDouaji/detr-doc-table-detection already pass all 3 EPs.

Acceptance Criteria

• microsoft/table-transformer-detection passes all 3 EPs (investigate VitisAI failure)
• microsoft/table-transformer-structure-recognition-v1.1-all passes all 3 EPs (investigate VitisAI failure)
• PekingU/rtdetr_r101vd_coco_o365 passes all 3 EPs (investigate VitisAI failure)
• PekingU/rtdetr_v2_r18vd passes all 3 EPs (investigate VitisAI failure)
• PekingU/rtdetr_r50vd_coco_o365 passes all 3 EPs (investigate QNN + VitisAI failure)