Research map / Label-efficient and robust learning / Recognition and long-tail
CSPCL: Category Semantic Prior Contrastive Learning for Deformable DETR-Based Prohibited Item Detectors
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CSPCL, a plug-and-play contrastive mechanism that aligns Deformable DETR decoder content queries with classifier-weight class prototypes via a CSP loss (truncated intra-class attraction plus adaptive inter-class repulsion) to improve X-ray prohibited item detection.
Problem
Overlapping objects in X-ray images couple foreground and background features, so general detectors perform poorly on prohibited item detection; in DETR-style decoders, content queries lose classification semantics over layers.
Approach
Use classifier weights as class prototypes to contrastively align content queries per category, with an intra-class gradient truncation (ITA) that stops over-homogenizing queries and an inter-class repulsion (IAR) scaled by prototype-prototype similarity, which outperforms N-pair and InfoNCE losses.
Key results
- On PIXray, CSPCL raises DINO AP from 64.3 to 67.5 (+3.2) and Deformable DETR AP from 36.6 to 41.8 (+5.2), with no added parameters or FPS cost.
- CSP loss outperforms N-pair (66.0) and InfoNCE (65.4) when plugged into DINO on PIXray.
- C-AO-DETR (AO-DETR + CSPCL, Swin-L, 512x512) reaches 70.8 overall AP on PIDray, the highest among compared models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CSPCL to Deformable DETR | Deformable DETR · PIXray | AP: 36.6 -> 41.8 (+5.2) | Helped |
| Add CSPCL to DINO | DINO · PIXray | AP: 64.3 -> 67.5 (+3.2) | Helped |
| Add CSPCL to AO-DETR | AO-DETR · OPIXray | mAP: 77.7 -> 78.5 (+0.8) | Helped |
| Add CSPCL to AO-DETR | AO-DETR · PIDray overall | AP: 70.0 -> 70.8 | Helped |
| Add CSPCL to AO-DETR | AO-DETR · CLCXray | AP | Helped |
| Replace CSP loss with N-pair loss | DINO + CSPCL · PIXray | AP: 64.3 -> 66.0 (+1.7) | Helped |
| Replace CSP loss with InfoNCE | DINO + CSPCL · PIXray | AP: 64.3 -> 65.4 (+1.1) | Helped |
| Add CSP loss instead of using no contrastive loss (ablation) | DINO · PIXray | AP: 64.3 -> 66.7 (+2.4) | Helped |
| Use ITA alone with gamma=5×10⁻³ (ablation) | DINO + CSPCL · PIXray | AP: 64.3 -> 65.4 (+1.1) | Helped |
| Use IAR alone with tau=0.3 (ablation) | DINO + CSPCL · PIXray | AP: 64.3 -> 65.6 (+1.3) | Helped |
| Apply CSPCL only at decoder layer 0 (ablation) | DINO + CSPCL · PIXray | AP: 64.3 -> 66.7 (+2.4) | Helped |
| Apply CSPCL only at decoder layer 5 (ablation) | DINO + CSPCL · PIXray | AP: 64.3 -> 60.6 (-3.7) | Hurt |
Surprising findings
- Aligning only shallow decoder layers performs better than aligning a single deep layer; aligning layer 5 alone drops AP by 3.7 versus baseline.
- On natural-image COCO, CSPCL gives only about 0.1% AP improvement, per the authors' limitations section.
Limitations
- Gains are small on natural-image data (about 0.1% AP on COCO).
- No error bars or statistical significance reported.
- Contrastive loss design relies on the classifier weights as prototypes and on the overlap phenomenon specific to X-ray images.
Experimental setup
Datasets: PIXray OPIXray PIDray CLCXray
Benchmarks: PIXray OPIXray PIDray CLCXray
Metrics: AP AP50 AP75 mAP FPS PARAMs GFLOPs
Baselines: Deformable DETR RT-DETR DINO AO-DETR N-pair loss InfoNCE loss Faster R-CNN Mask R-CNN ForkNet DOAM LAreg Cascade R-CNN
Builds on: Deformable DETR DINO MMCL AO-DETR InfoNCE
Closest papers by meaning
- SIXray: A Large-Scale Security Inspection X-Ray Benchmark for Prohibited Item Discovery in Overlapping Images (CVPR 2019 · similarity 0.87)
- Can a Second-View Image Be a Language? Geometric and Semantic Cross-Modal Reasoning for X-ray Prohibited Item Detection (CVPR 2026 · similarity 0.85)
- DetKDS: Knowledge Distillation Search for Object Detectors (ICML 2024 · similarity 0.82)
- Dual-view X-ray Detection: Can AI Detect Prohibited Items from Dual-view X-ray Images like Humans? (CVPR 2025 · similarity 0.82)
- M²RL-Net: Multi-View and Multi-Level Relation Learning Network for Weakly-Supervised Image Forgery Detection (AAAI 2025 · similarity 0.81)
- A Bias-Free Training Paradigm for More General AI-generated Image Detection (CVPR 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.