What are the three loss functions used to train the reconstruction module in the proposed method, and what is the purpose of each?

The three loss functions are decision-level loss, feature-level loss, and L2 loss. Decision-level loss combines score loss and drift loss to align the tracker's classification and regression outputs; feature-level loss reduces the distance between feature maps of reconstructed and original frames; L2 loss minimizes pixel-wise difference between reconstructed and original frames.

The reconstruction module is trained with three newly designed loss functions: decision-level loss, feature-level loss, and L2 loss. The decision-level loss is split into score loss and drift loss. Score loss reduces the difference between the target and background confidence scores produced from the original clean frame and the reconstructed frame. Drift loss makes the predicted bounding box size and center position in the reconstructed frame close to those in the clean frame. Feature-level loss uses the feature maps extracted from the tracker's backbone and minimizes the sum of l2 norm distances between feature maps from the original frame and the reconstructed frame across selected layers. L2 loss minimizes the pixel-wise distance between the clean original frame and the reconstructed frame, encouraging overall visual similarity.