New COD-TDQ method boosts quantized Transformer performance for object detection

Researchers have developed a new method called COD-TDQ to improve the performance of Transformer-based models for camouflaged object detection (COD) when using aggressive post-training W4A4 quantization. They identified that heavy-tailed background tokens in COD tasks inflate the activation range, causing crucial boundary cues to be lost. COD-TDQ addresses this by using Direct-Sum Token-Group to suppress range domination and Dual-Constraint Range Projection to maintain bounded step-to-dispersion ratios and zero-bin mass. This approach consistently achieves significantly higher scores on COD benchmarks compared to existing quantization methods without requiring retraining. AI

IMPACT This research could enable more efficient deployment of computer vision models on resource-constrained devices.

RANK_REASON The item is a research paper detailing a new method for improving model quantization for a specific computer vision task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• Camouflaged object detection via boundary refinement
• CFRN
• COD-TDQ
• Direct-Sum Token-Group
• Dual-Constraint Range Projection
• ESCNet: Entity-enhanced and Stance Checking Network for Multi-modal Fact-Checking
• Tianqi Li
• Transformer++
• W4A4

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →