.png){kind=icon}
.png){kind=icon}
Return
AMETracker: transformer tracking with asymmetric mix encoder
.png){kind=icon}
PRE
AI
C
张
Y
DOI：10.1007/s11760-026-05273-x.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Feature fusion is a critical step in visual object tracking. However, previous trackers have mixed both the template features and the search region features when performing fusion, resulting in some irrelevant features in the background of the search region being fused into the template features. This contamination of the template features has had a detrimental effect on the tracker's ability to accurately locate the target. Additionally, the feature fusion process at the decoder stage has also increased the probability of positional inaccuracy. In order to address these issues, we propose an asymmetric mix encoder and a symmetric cross decoder for improved feature enhancement and fusion. The encoder structure efficiently preserves the integrity of the template features, prevents any disruption caused by irrelevant features in the search region, and consistently delivers accurate template feature comparisons throughout the tracking procedure. The decoder structure makes full use of virgin template features and search region features for deep fusion. Furthermore, we design a pixel-level reconstruction module to rebuild the image after feature fusion, which helps to obtain more accurate position information of the target, thus aiding our tracker to locate the target more accurately. The effectiveness of AMETracker in preventing template contamination and its efficient fusion strategy helped it achieve state of the art performance on seven datasets.
Keywords:
Visual tracking
Transformer
Asymmetric mix encoder
Symmetric cross decoder
Pixel-level reconstruction module
Journal
IF:
2.1
Papers:
778
Citations:
4.6K
Organization
Cited Papers