Foreground Information Guidance for Siamese Visual Tracking

Abstract

Existing Siamese network based trackers are easily disturbed by large deformation, occlusion and distractor objects in the background. By comparing these trackers, we observe that the monotonous positive pairs usually have limited challenging factors (Occlusion, Deformation, etc.), which may make the learned features less robust. In addition, the foreground information of the substantial training data is utilized directly without deeper exploration. Thus, the trackers cannot effectively discriminate the foreground from the semantic backgrounds. In this paper, we focus on modifying the Siamese tracker by enriching the positive pairs and taking further advantage of the foreground information. During the offline training phase, a simple sampling strategy is adopted to enrich the challenging factors in positive pairs, which can effectively enhance the robustness of the tracker. At the same time, we highlight the foreground information by padding the background, and the information is utilized to generate a novel padding loss, which guides the tracker to pay less attention to the distractors in the background. Moreover, an improved feature information fusion is adopted to update the template, so that the tracker can adapt to the drastic appearance changes. Comprehensive experiments on the OTB and the VOT benchmarks demonstrate that our proposed tracker can achieve outstanding performance in both accuracy and robustness. INDEX TERMS Visual tracking, Siamese network, foreground information, feature information fusion.

Siamese Visual Object Tracking: A Survey

“…The logistic loss was adopted in trackers SiamFC [12], Correlation filter network (CFNet) [54] and Dynamic Siamese network (DSiam) [55]. This loss was a foundation to many upcoming trackers, e.g., Siamese classification and regression networks (SiamCAR) [20] or Foreground information guidance for Siamese visual tracking (FIGSiam) [56], where they extended this loss by adding more terms.…”

“…This analysis opened up the possibility to use a broader set of backbones, and successive works (e.g., [56]) also adopted the spatial-aware sampling strategy to counterbalance the breaking of strict translation invariance. An analysis by Han et al [67] also confirmed that using padding in Siamese backbones had a huge negative impact on tracking performance.…”

“…The context itself may pollute the response map if similar interference is present. In such a case, Li et al [56] proposed a foreground information guidance module for their FIGSiam tracker. The core idea was based on a padding image (derived from the search image) that contained the target object but the remaining context (outside the BBOX region) was filled with a mean color computed independently for all three channels.…”

“…One obvious approach to simulate occlusion is to cover a specific part of the object region in the image. For example, Li et al [56] added occlusion as part of data augmentation for training their FIGSiam architecture. Let w and h be the width and height of the ground-truth BBOX denoting the location of the target object.…”

“…Nowadays, the two most common ways to achieve template updating are either linear interpolation or multi-template updating [92]. As for Siamese trackers, the constant template strategy (i.e., the template is initialized in the first frame and never changed) that is used in this type of trackers exacerbates their ability to adapt to the drastic appearance changes [56]. Trackers that exploit classification and online model updating perform among the best in terms of accuracy, but they are the slowest ones (even 1-2 FPS).…”