Research map / Video / Action recognition
AssembleNet++: Assembling Modality Representations via Attention Connections - Supplementary Material -
Michael S. Ryoo, AJ Piergiovanni, Juhana Kangaspunta, Anelia Angelova
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 omnipresent object-modality connections (ablation) | AssembleNet++ block connectivity · Charades v1 | mAP: 47.18 -> 50.43 | Improved |
| replace weighted connections with peer-attention (no object) (ablation) | AssembleNet++ · Charades v1 | mAP: 47.18 -> 52.39 | Improved |
| replace weighted connections with peer-attention (with object) (ablation) | AssembleNet++ · Charades v1 | mAP: 50.43 -> 56.38 | Improved |
| use peer-attention instead of self-attention (ablation) | AssembleNet++ with object · Charades v1 | mAP: 55.40 -> 56.38 | Improved |
| use static attention instead of none (ablation) | AssembleNet++ with object · Charades v1 | mAP: 50.43 -> 51.15 | Improved |
| add peer-attention instead of 1x1 conv with same parameters (ablation) | AssembleNet++ · Charades v1 | mAP: 50.24 -> 56.38 | Improved |
| use searched peer-attention instead of random connectivity (ablation) | AssembleNet++ · Charades v1 | mAP: 53.40 -> 56.38 | Improved |
| use searched peer-attention instead of random connectivity (ablation) | AssembleNet++ · Toyota Smarthome | mean per-class accuracy: 60.23 -> 63.64 | Improved |
| add 1x1 conv instead of attention (ablation) | AssembleNet++ base · Toyota Smarthome | mean per-class accuracy: 59.16 -> 59.44 | No clear effect |
| add object modality and peer-attention (ablation) | AssembleNet-50 baseline · Toyota Smarthome | mean per-class accuracy: 57.42 -> 63.64 | Improved |
| add self-attention only (no object) (ablation) | AssembleNet-50 baseline · Toyota Smarthome | classification accuracy: 77.59 -> 77.77 | No clear effect |
| add object modality and attention to RGB R(2+1)D (ablation) | RGB R(2+1)D · Charades v1 | mAP: 36.51 -> 45.30 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AssembleNet-50 — AssembleNet: Searching for Multi-Stream Neural Connectivity in Video Architectures (ICLR 2020) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Attentional Pooling for Action Recognition (NeurIPS 2017 · similarity 0.87)
- AssembleNet: Searching for Multi-Stream Neural Connectivity in Video Architectures (ICLR 2020 · similarity 0.87)
- AOE-Net: Entities Interactions Modeling with Adaptive Attention Mechanism for Temporal Action Proposals Generation (IJCV 2022 · similarity 0.87)
- Learning what and where to attend (ICLR 2019 · similarity 0.86)
- Class Semantics-Based Attention for Action Detection (ICCV 2021 · similarity 0.86)
- A^2-Nets: Double Attention Networks (NeurIPS 2018 · 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.
