STransLOT: splitting-refusion transformer for low-light object tracking

In the field of tracking, more and more trackers are using the great potential of the transformer to form the framework. Most of them use the Siamese-based backbone and employ the attention mechanism to capture the spatio-temporal features, which benefits the similarity learning and establishing the positional relationship between the template patch and the search region. However, tracking a target accurately in low-light scenarios is one of the most challenging tasks in recent years. To alleviate this defect, we propose an improved Splitting-refusion Transformer for Low-light Object Tracking (STransLOT). Building on the irreplaceable success that Transformer trackers have achieved in visual tracking this year, our STransLOT is combined with a Transformer-like feature fusion module and a classical prediction head. The pixel-level splitting module splits the original image into the part high-light image and part low-light image, while the refusion module fuses the feature maps of these three inputs to improve the low-light feature representation. Experiments show that our STransLOT achieves remarkable results on the LOTD50 dataset and other low-light sequences of public benchmarks.