CCLT: Correction-Conditioned Low-Rank Exchange for Parameter-Efficient RGB-X Tracking

Abstract Parameter-efficient fine-tuning (PEFT) offers a practical way to adapt pretrained RGB trackers to RGB-X tracking, where paired multimodal training data are limited. Despite their architectural diversity, existing PEFT approaches typically fuse the RGB and X branches through two parallel directional corrections. The (X→RGB) correction updates the pre-fusion RGB state using the pre-fusion X state, while the (RGB→X) correction updates the pre-fusion X state using the pre-fusion RGB state. This parallel formulation prevents the (RGB→X) correction from using the corrected RGB state produced by the (X→RGB) correction. However, the source-state dependency between the two directional corrections remains underexplored. To model this source-state dependency under a compact trainable budget, we introduce Correction-Conditioned Low-rank Tracking (CCLT), a parameter-efficient RGB-X tracker that combines sparsely inserted bidirectional Lift exchange (BiLift) modules with lightweight Intra-Stream LoRA (IntraLoRA) adapters. BiLift applies an (X→RGB→X) corrective relay at one selected backbone depth and an (RGB→X→RGB) relay at another, allowing the second directional correction at each depth to read a freshly corrected source state. IntraLoRA reuses low-rank key/value factors across the RGB and X streams to adapt within-stream representations while keeping the shared backbone frozen, thereby balancing cross-modal exchange with parameter-efficient representation adaptation. Together, these innovations enable CCLT to achieve a more favorable balance between tracking performance and computational efficiency than existing state-of-the-art methods across RGB-T, RGB-D and RGB-E tracking tasks.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11441577/v1
Primary Topic
Video Surveillance and Tracking Methods
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

CCLT: Correction-Conditioned Low-Rank Exchange for Parameter-Efficient RGB-X Tracking

Qing Song, Xingyu Tan, Chun Liu, Mengjie Hu
4 links
Research Square
Video Surveillance and Tracking Methods
preprint

CCLT: Correction-Conditioned Low-Rank Exchange for Parameter-Efficient RGB-X Tracking

Qing Song, Xingyu Tan, Chun Liu, Mengjie Hu
preprint en

Abstract

Abstract Parameter-efficient fine-tuning (PEFT) offers a practical way to adapt pretrained RGB trackers to RGB-X tracking, where paired multimodal training data are limited. Despite their architectural diversity, existing PEFT approaches typically fuse the RGB and X branches through two parallel directional corrections. The (X→RGB) correction updates the pre-fusion RGB state using the pre-fusion X state, while the (RGB→X) correction updates the pre-fusion X state using the pre-fusion RGB state. This parallel formulation prevents the (RGB→X) correction from using the corrected RGB state produced by the (X→RGB) correction. However, the source-state dependency between the two directional corrections remains underexplored. To model this source-state dependency under a compact trainable budget, we introduce Correction-Conditioned Low-rank Tracking (CCLT), a parameter-efficient RGB-X tracker that combines sparsely inserted bidirectional Lift exchange (BiLift) modules with lightweight Intra-Stream LoRA (IntraLoRA) adapters. BiLift applies an (X→RGB→X) corrective relay at one selected backbone depth and an (RGB→X→RGB) relay at another, allowing the second directional correction at each depth to read a freshly corrected source state. IntraLoRA reuses low-rank key/value factors across the RGB and X streams to adapt within-stream representations while keeping the shared backbone frozen, thereby balancing cross-modal exchange with parameter-efficient representation adaptation. Together, these innovations enable CCLT to achieve a more favorable balance between tracking performance and computational efficiency than existing state-of-the-art methods across RGB-T, RGB-D and RGB-E tracking tasks.

Research Square
Beijing University of Posts and Telecommunications (CN)
Video Surveillance and Tracking Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon