Research map / Video / Video segmentation and tracking
Efficient Image Pre-Training with Siamese Cropped Masked Autoencoders
Alexandre Eymaël, Renaud Vandeghen, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, Marc Van Droogenbroeck
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 two different crops (global-to-local) instead of the same view for both inputs (ablation) | CropMAE · DAVIS-2017 (IN Subset) | J&F_m: 36.6 -> 60.4 | Improved |
| global-to-local crops instead of local-to-global (ablation) | CropMAE · DAVIS-2017 (IN Subset) | J&F_m: 55.9 -> 60.4 | Improved |
| global-to-local crops instead of independent random crops (ablation) | CropMAE · DAVIS-2017 (IN Subset) | J&F_m: 60.0 -> 60.4 | No clear effect |
| raise masking ratio from 0.75 to 0.985 (ablation) | CropMAE · DAVIS-2017 (IN Subset) | J&F_m: 45.3 -> 60.4 | Improved |
| raise masking ratio from 0.985 to 0.99 (one visible patch) (ablation) | CropMAE · DAVIS-2017 (IN Subset) | J&F_m: 60.4 -> 58.6 | Hurt |
| deeper decoder (8 blocks vs 4) (ablation) | CropMAE · DAVIS-2017 (IN Subset) | J&F_m: 60.4 -> 57.0 | Hurt |
| add color jitter augmentation (ablation) | CropMAE · DAVIS-2017 (IN Subset) | J&F_m: 60.3 -> 56.2 | Hurt |
| add random horizontal flip (ablation) | CropMAE · DAVIS-2017 (IN Subset) | J&F_m: 60.3 -> 60.4 | No clear effect |
| pre-train on ImageNet subset images instead of K400 video frames | CropMAE · DAVIS-2017 | J&F_m: 58.6 -> 60.4 | Improved |
| CropMAE image-based pretext vs SiamMAE video frames, 400 epochs | CropMAE · JHMDB | PCK@0.1: 46.1 -> 43.6 | Hurt |
| CropMAE on images vs SiamMAE on K400, training time | CropMAE · pre-training speedup | speedup: x1.0 -> x23.8 | Improved |
Closest papers by meaning
- Siamese Masked Autoencoders (NeurIPS 2023 · similarity 0.91)
- VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training (NeurIPS 2022 · similarity 0.89)
- VideoMAE V2: Scaling Video Masked Autoencoders With Dual Masking (CVPR 2023 · similarity 0.89)
- Masked Autoencoders As Spatiotemporal Learners (NeurIPS 2022 · similarity 0.88)
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · similarity 0.86)
- Asymmetric Masked Distillation for Pre-Training Small Foundation Models (CVPR 2024 · similarity 0.86)
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.
