Research map / Detection and segmentation / Open-set and few-shot detection
Invisible Shortcuts: Why Vision Encoders Know Your Camera
Vladan Stojnic, Ryan Ramos, Giorgos Kordopatis-Zilos, Noa Garcia, Giorgos Tolias
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 |
|---|---|---|---|
| use ImageNet21k instead of ImageNet1k pretraining (stronger metadata-semantics correlation) | timm models · IN1k/IN21k diagnostics | metadata sensitivity (MP, SPD): Cramer's V 0.047 -> 0.067 | Hurt |
| increase JPEG-semantics correlation p_i during training (ablation) | ResNet50 · IN1k diagnostics | SPD_p and MP | Hurt |
| increase JPEG-semantics correlation p_i during training (ablation) | ResNet50 · IN1k | SP accuracy | Hurt |
| use stronger acquisition-topic correlation in Re-LAION-2B subset (ablation) | ResNet50 CLIP · Re-LAION-2B diagnostics | SPD: Cramer's V 0.255 -> 0.396 | Hurt |
| add color jitter, grayscale and blur augmentations (ablation) | ResNet50 · IN1k | SPD_p JPEG: 2.7 -> 1.4 | Improved |
| add color jitter, grayscale and blur augmentations (ablation) | ResNet50 · IN1k | SP accuracy: 78.2 -> 77.3 | No clear effect |
| add post-hoc adversarial linear layer (ablation) | ResNet50 · IN1k | SPD_p JPEG: 2.7 -> 0.8 | Improved |
| combine augmentations and post-hoc mitigation (ablation) | ResNet50 · IN1k | SPD_p JPEG: 2.7 -> 0.3 | Improved |
| combine augmentations and post-hoc mitigation (ablation) | ResNet50 · IN1k | SP accuracy: 78.2 -> 76.5 | Hurt |
| add Gaussian blur augmentation (ablation) | ResNet50 · IN1k | processing vs acquisition sensitivity | Mixed |
| train with strong JPEG-semantics correlation (ablation) | ResNet50 · Generated image detection | avg accuracy: 57.3 -> 60.4 | Improved |
| apply post-hoc metadata mitigation | foundation models · ImageNet-C/R/Sketch | accuracy | Mixed |
Closest papers by meaning
- Perception Encoder: The best visual embeddings are not at the output of the network (NeurIPS 2025 · similarity 0.87)
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · similarity 0.87)
- Beyond Encoder Accumulation: Measuring Encoder Roles in Multi-Encoder VLMs (NeurIPS 2026 · similarity 0.86)
- Enhancing Fine-Grained Vision-Language Pretraining with Negative Augmented Samples (AAAI 2025 · similarity 0.85)
- Representation-Level Counterfactual Calibration for Debiased Zero-Shot Recognition (NeurIPS 2025 · similarity 0.85)
- CLIP Tricks You: Training-free Token Pruning for Efficient Pixel Grounding in Large Vision-Language Models (ICML 2026 · similarity 0.85)
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.
