Research map / Vision-language and multimodal / Cross-modal learning
IRGPT: Understanding Real-world Infrared Image with Bi-cross-modal Curriculum on Large-scale Benchmark
Zhe Cao, Jin Zhang, Ruiheng Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add bi-cross-modal curriculum pre-training vs random order (ablation) | IRGPT · IR-TD zero-shot | psum: 307.18 -> 328.65 | Improved |
| add bi-cross-modal curriculum pre-training vs random order (ablation) | IRGPT · IR-TD zero-shot | nsum: 97.49 -> 93.87 | Improved |
| use easy-to-hard curriculum instead of hard-to-easy (ablation) | IRGPT · IR-TD fine-tuned | psum: 433.24 -> 485.79 | Improved |
| use curriculum instead of random order (ablation) | IRGPT · IR-TD fine-tuned | nsum: 12.94 -> 4.39 | Improved |
| add curriculum-pretrained IRGPT to InternVL2-8B baseline | InternVL2-8B · IR-TD zero-shot | psum: 252.30 -> 328.65 | Improved |
| add curriculum-pretrained IRGPT to InternVL2-8B baseline | InternVL2-8B · IR-TD zero-shot | nsum: 125.43 -> 93.87 | Improved |
| fine-tuned IRGPT (8B) vs larger InternVL2-26B | InternVL2-26B · IR-TD fine-tuned | psum: 473.78 -> 485.79 | Improved |
| random pre-training beats curriculum on pedestrian counting | IRGPT · IR-TD zero-shot | Pedestrian Counting MAE: 47.17 vs 47.30 | Hurt |
| add loss variation rate weight alpha (ablation) | IRGPT (L1+L2) · IR-TD fine-tuned | psum: 481.36 -> 485.79 | Improved |
| add loss variation rate weight alpha (ablation) | IRGPT (L1+L2) · IR-TD fine-tuned | nsum: 7.08 -> 4.39 | Improved |
| add Lesson 2 (IR-T) to Lesson 1 curriculum (ablation) | IRGPT · IR-TD fine-tuned | psum: 466.81 -> 485.79 | Improved |
| add Lesson 1 (IR-VIS) to Lesson 2 curriculum (ablation) | IRGPT · IR-TD fine-tuned | psum: 470.23 -> 485.79 | Improved |
Closest papers by meaning
- IF-Bench: Benchmarking and Enhancing MLLMs for Infrared Images with Generative Visual Prompting (CVPR 2026 · similarity 0.89)
- SynthRGB-T: Language-Vision Guided Image Translation for Diversity Synthesis (CVPR 2026 · similarity 0.86)
- UNIP: Rethinking Pre-trained Attention Patterns for Infrared Semantic Segmentation (ICLR 2025 · similarity 0.85)
- WikiDO: A New Benchmark Evaluating Cross-Modal Retrieval for Vision-Language Models (NeurIPS 2024 · similarity 0.85)
- Quality-Driven Curation of Remote Sensing Vision-Language Data via Learned Scoring Models (NeurIPS 2025 · similarity 0.84)
- Rethinking Infrared Small Target Detection: A Foundation Driven Efficient Paradigm (NeurIPS 2026 · similarity 0.84)
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.
