AI Paper Atlas

Research map / Vision-language and multimodal / Robot action and navigation

Multimodal Transformer with Variable-Length Memory for Vision-and-Language Navigation

Chuang Lin, Yi Jiang, Jianfei Cai, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan

ECCV 2022 · Vision-language and multimodal › Robot action and navigation · approach: Reasoning-enhanced vision-language navigation · 37 citations (OpenAlex)

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 changedWhereMeasuredEffect
add the memory-aware consistency loss to MTVM (ablation)MTVM · R2R val unseenSR / SPL: 64.0 -> 65.7 SR; 58.6 -> 59.4 SPLImproved
use random word dropping as data augmentation instead of the consistency loss (ablation)MTVM · R2R val unseenSR / SPL: 64.0 -> 64.5 SR; 58.6 -> 57.8 SPLMixed
replace the recurrent state history with the cross-modal activation memory bankVLN-BERT · R2R val unseenSR / SPL: 63 -> 65.7 SR; 57 -> 59.4 SPLImproved
use cross-modal memory instead of hierarchical visual-only history (HAMT)HAMT · R2R val unseenSR / SPL: 63.5 -> 65.7 SR; 57.5 -> 59.4 SPLImproved
use memory with single-direction cross-attention (MTVM-dagger*) instead of VLN-BERT's hidden-state networkVLN-BERT · R2R val unseenSR / SPL / memory: 63.3 -> 63.6 SR; 57.5 -> 58.2 SPL; 8.6GB -> 8.4GBImproved
use bidirectional cross-attention instead of language-only keys/values in MTVM (ablation)MTVM without consistency loss · R2R val unseenSR / SPL / memory: 63.6 -> 64.0 SR; 58.2 -> 58.6 SPL; 8.4GB -> 17.9GBMixed
increase the memory bank size / use variable-length memory storing every step (ablation)MTVM · R2R val seen and unseenNE, SR, SPLImproved
raise the word dropping rate from 0.1 to 0.5 (ablation)MTVM · R2R val seen and unseenSR, SPLImproved
raise the word dropping rate from 0.5 to 0.7 (ablation)MTVM · R2R val seen and unseenSR, SPLHurt
vary consistency loss weights lambda_s and lambda_m in the range 2 to 8 (ablation)MTVM · R2R val unseenSPLNo clear effect
use cross-modal history (VLN-BERT recurrent state) instead of visual-only oriented observations (E.T.)history encoding · R2R val unseenSR / SPL: 59.0 -> 63 SR; 54.5 -> 57 SPLImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy