Research map / Detection and segmentation / Open-set and few-shot detection
ATAS: Any-to-Any Self-Distillation for Enhanced Open-Vocabulary Dense Prediction
Juan Yeo, Soonwoo Cha, Jiwoo Song, Hyunbin Jin, Taesup Kim
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 Global-to-Local Distillation to plain CLIP (ablation) | ATAS · F-ViT detection | AP50: 31.40 -> 45.98 | Improved |
| add Local-to-Local Distillation (ablation) | ATAS · Detection AP50 | AP50: 45.98 -> 45.39 | Hurt |
| add Global-to-Global Distillation (ablation) | ATAS · MaskCLIP segmentation | mIoU: 37.89 -> 39.34 | Improved |
| use contrastive objective instead of cosine in GLD (ablation) | ATAS GLD · MaskCLIP segmentation | mIoU: 36.98 -> 37.62 | Improved |
| add selective patch weighting with contrastive GLD (ablation) | ATAS GLD · Segmentation and detection | mIoU / AP50: 37.62 -> 37.6 mIoU; 45.88 -> 45.98 AP50 | No clear effect |
| remove GGD loss (ablation) | ATAS student · ImageNet accuracy | accuracy: 40.92 -> 67.25 | Improved |
| increase LLD weight from 0.01 to 1 (ablation) | ATAS · Segmentation vs detection | mIoU / AP50: seg 39.11 -> 42.31, det 45.87 -> 44.57 (with GGD 0.01 vs 1) | Mixed |
| set contrastive temperature to 1 versus 0.1 or 10 (ablation) | ATAS · Detection AP50 | AP50: 45.11 -> 46.38 (tau 0.1 -> 1) | Improved |
| train CLIP with ATAS instead of CLIPSelf for zero-shot MaskCLIP | CLIP ViT-B · Zero-shot segmentation (MaskCLIP) | average mIoU: 26.4 -> 31.2 | Improved |
| replace CLIP weights with ATAS in CAT-Seg v2 fine-tuning | CAT-Seg v2 · Fine-tuned segmentation average | mIoU: 48.6 -> 49.1 | Improved |
| replace CLIPSelf weights in CAT-Seg v2 | CAT-Seg v2 · Fine-tuned segmentation average | mIoU: 48.6 -> 48.3 | Hurt |
| ATAS versus CLIPSelf in F-ViT detection | F-ViT · OV-COCO | AP50 novel: 33.6 -> 37.2 | Improved |
Closest papers by meaning
- EdaDet: Open-Vocabulary Object Detection Using Early Dense Alignment (ICCV 2023 · similarity 0.90)
- Reconstructing CLIP for Open-Vocabulary Dense Perception (CVPR 2026 · similarity 0.90)
- DINOv2 Meets Text: A Unified Framework for Image- and Pixel-Level Vision-Language Alignment (CVPR 2025 · similarity 0.89)
- Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models (CVPR 2024 · similarity 0.89)
- Unlocking Fine-Grained Perception in CLIP via Structurally-Aware Latent Masked Modeling (NeurIPS 2026 · similarity 0.89)
- CLIPSelf: Vision Transformer Distills Itself for Open-Vocabulary Dense Prediction (ICLR 2024 · similarity 0.88)
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.
