Research map / Video / Video segmentation and tracking
DropMAE: Masked Autoencoders With Spatial-Attention Dropout for Tracking Tasks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 adaptive spatial-attention dropout to video MAE reconstruction (ablation) | TwinMAE pre-training · GOT-10k | AO: 72.9 -> 74.8 | Helped |
| add adaptive spatial-attention dropout to video MAE reconstruction (ablation) | TwinMAE pre-training · DAVIS-17 | J &F: 80.7 -> 82.7 | Helped |
| use DropMAE video pre-training instead of ImageNet MAE pre-training | ViT-B/16 downstream tracker · GOT-10k | AO: 73.7 -> 74.8 | Helped |
| use DropMAE video pre-training instead of ImageNet MAE pre-training | ViT-B/16 downstream segmenter · DAVIS-17 | J &F: 81.7 -> 82.7 | Helped |
| increase DropMAE training from 400 to 800 epochs (ablation) | DropMAE, K400 · GOT-10k | AO: 73.2 -> 74.8 | Helped |
| pre-train on K700 rather than K400 (ablation) | DropMAE · GOT-10k | AO: 74.8 -> 75.9 | Helped |
| increase spatial-attention dropout ratio from 0.05 to 0.1 (ablation) | DropMAE pre-training · DAVIS-17 | J&F: 80.3 -> 81.3 | Helped |
| increase spatial-attention dropout ratio from 0.1 to 0.2 (ablation) | DropMAE pre-training · DAVIS-17 | J&F: 81.3 -> 79.8 | Hurt |
| apply ASAD in the encoder as well as the decoder (ablation) | DropMAE-K400-400E · GOT-10k | AO: 73.2 -> 73.1 | Hurt |
| add domain-specific tracking videos to K400 pre-training (ablation) | DropMAE-K400-400E · GOT-10k | AO: 73.2 -> 73.4 | Helped |
| remove frame identity embeddings (ablation) | DropMAE-K400-400E · GOT-10k | AO: 73.2 -> 72.9 | Hurt |
| use more action-diverse K700 videos rather than K400 videos (ablation) | DropMAE pre-training · GOT-10k | AO: 73.2 -> 75.6 | Helped |
Closest papers by meaning
- TrackMAE: Video Representation Learning via Track Mask and Predict (CVPR 2026 · similarity 0.88)
- VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training (NeurIPS 2022 · similarity 0.88)
- VideoMAC: Video Masked Autoencoders Meet ConvNets (CVPR 2024 · similarity 0.88)
- Video Diffusion Models Excel at Tracking Similar-Looking Objects Without Supervision (NeurIPS 2025 · similarity 0.86)
- Extending Video Masked Autoencoders to 128 frames (NeurIPS 2024 · similarity 0.86)
- Emergent Temporal Correspondences from Video Diffusion Transformers (NeurIPS 2025 · 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.