AI Paper Atlas

Research map / Detection and segmentation / Tracking

Tracking by Natural Language Specification with Long Short-term Context Decoupling

Ding Ma, Xiangqian Wu

ICCV 2023 · Detection and segmentation › Tracking · approach: Transformer-based video object tracking · 27 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 both SCM and LCP branches with LSM module to frame-token baseline (ablation)DecoupleTNL · TNL2k (NL + Box)AUC: 0.318 -> 0.567Improved
add short-term context-matching (SCM) branch (ablation)frame-token baseline · LaSOT (NL + Box)AUC: 0.427 -> 0.605Improved
add long-term context-perceiving (LCP) branch (ablation)frame-token baseline · LaSOT (NL + Box)AUC: 0.427 -> 0.620Improved
use LSM module to fuse long and short-term tokens instead of sum, concat or multi-conv (ablation)DecoupleTNL · TNL2k (NL + Box)AUC: 52.9 -> 56.7 (vs sum)Improved
use short-term temporal scope T=5 rather than 2, 3, 4 or 6 (ablation)SCM branch · TNL2k (NL + Box)AUC: 56.7 vs 56.4, 55.2, 53.8, 56.1Mixed
use sigmoid or sigmoid+multi-conv for frame tokens instead of scale-and-shift modulation (ablation)LSM module · TNL2k (NL + Box)AUC: 56.7 -> 55.1 (sigmoid), 55.6 (sigmoid+multi-conv)Hurt
stack 4 LSM modules versus 1 to 5 (ablation)LSM module · TNL2k (NL + Box)AUC: 56.7 at 4 stacks; 55.7 to 56.6 for othersNo clear effect
use contrastive (InfoNCE) learning for the context branches instead of token concatenation (ablation)DecoupleTNL · TNL2k (NL + Box)AUC and training timeImproved
add long short-term context decoupling strategy to existing TNL trackers (TNL, SNLT, CapsuleTNL, AdaSwitcher)existing TNL trackers · TNL2kAUCImproved
give baseline trackers the same ResNet50 + transformer backbone (ablation)ToMP, TransT, AiATrack, TrDiMP, SNLT, AdaSwitcher · OTB-lang and LaSOTAUCNo clear effect
initialize with NL only instead of NL + Box (ablation)DecoupleTNL · TNL2kAUC: 0.567 -> 0.407Hurt

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