AI Paper Atlas

Research map / Video / Video segmentation and tracking

TDViT: Temporal Dilated Video Transformer for Dense Video Tasks

Guanxiong Sun, Yang Hua, Guosheng Hu, Neil Robertson

ECCV 2022 · Video › Video segmentation and tracking · approach: Video semantic and panoptic segmentation · 2 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
replace Swin-T backbone with TDViT-TFaster R-CNN · ImageNet VIDAP box: 47.1 -> 49.1Improved
replace Swin-S backbone with TDViT-SFaster R-CNN · ImageNet VIDAP box: 52.6 -> 55.4Improved
add advanced variant (two extra TDTBs in stage 3) to TDViT-TTDViT-T · ImageNet VIDAP box: 49.1 -> 50.9Improved
replace R-101 backbone with TDViT-T in SELSASELSA · ImageNet VIDAP50 box: 81.5 -> 83.9Improved
replace Swin-T with TDViT-TMaskTrack R-CNN · YouTube-VIS 2019AP mask: 33.7 -> 35.4Improved
replace Swin-S with TDViT-SMaskTrack R-CNN · YouTube-VIS 2019AP mask: 36.8 -> 37.7Improved
use split scheme rather than factorised scheme (ablation)TDViT-T · ImageNet VIDAP50 box: 78.0 -> 78.5Improved
add 2 more TDTBs in stage 3 (6 to 8 blocks) (ablation)TDViT-T · ImageNet VIDAP50 box: 78.5 -> 79.9Improved
increase temporal dilations from (1,2,4,8) to (4,8,16,32) (ablation)TDViT-T · ImageNet VIDAP50 box: 77.2 -> 78.5Improved
increase temporal dilations further to (5,10,20,40) (ablation)TDViT-T · ImageNet VIDAP50 box: 78.5 -> 78.4No clear effect
use patch shuffle sampling instead of temporal earliest (ablation)TDViT-T · ImageNet VIDAP50 box: 78.5 -> 78.8; FPS 23.9 -> 19.7Mixed
use channel shuffle sampling instead of temporal earliest (ablation)TDViT-T · ImageNet VIDAP50 box: 78.5 -> 77.7Hurt

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