Research map / Detection and segmentation / CNN features
MViTv2: Improved Multiscale Vision Transformers for Classification and Detection
Check 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 |
|---|---|---|---|
| Replace absolute positional embeddings with decomposed relative positional embeddings (ablation) | MViTv2-S · COCO val | APbox: 49.3 -> 49.9 | Helped |
| Add decomposed relative positional embeddings (ablation) | MViTv2-S · ImageNet-1K | accuracy: 83.5 -> 83.6 | No clear effect |
| Add absolute embeddings alongside decomposed relative embeddings (ablation) | MViTv2-S · ImageNet-1K | accuracy: 83.6 -> 83.7 | No clear effect |
| Add a residual pooling connection (ablation) | MViTv2-S · ImageNet-1K | accuracy: 83.3 -> 83.6 | Helped |
| Add a residual pooling connection (ablation) | MViTv2-S · COCO val | APbox: 48.5 -> 49.3 | Helped |
| Add full Q pooling without a residual connection (ablation) | MViTv2-S · ImageNet-1K | accuracy: 83.3 -> 83.1 | Hurt |
| Add full Q pooling together with the residual connection (ablation) | MViTv2-S · COCO val | APbox: 48.5 -> 49.9 | Helped |
| Add full Q pooling together with the residual connection (ablation) | MViTv2-S 16×4 · Kinetics-400 | top-1 accuracy: 79.3 -> 81.0 | Helped |
| Use decomposed space/time instead of joint space/time relative embeddings (ablation) | MViTv2-S 16×4 · Kinetics-400 | training throughput: 8.4 -> 66.6 clip/s | Helped |
| Pretrain on ImageNet-1K rather than train from scratch (ablation) | MViTv2-S 16×4 · Kinetics-400 | top-1 accuracy: 81.2 -> 82.2 | Helped |
| Pretrain on ImageNet-21K rather than train from scratch (ablation) | MViTv2-S 16×4 · Kinetics-400 | top-1 accuracy: 81.2 -> 82.6 | Helped |
| Increase pooling stride from 4 to 8 (ablation) | MViTv2-S · ImageNet-1K | accuracy: 83.6 -> 83.2 | Hurt |
Closest papers by meaning
- MSTDiff: Multiscale-Aware Transformer Diffusion Network for Video Object Detection (AAAI 2026 · similarity 0.88)
- MPViT: Multi-Path Vision Transformer for Dense Prediction (CVPR 2022 · similarity 0.87)
- ResT V2: Simpler, Faster and Stronger (NeurIPS 2022 · similarity 0.87)
- Multiscale Vision Transformers Meet Bipartite Matching for Efficient Single-stage Action Localization (CVPR 2024 · similarity 0.86)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.86)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · 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.