DomainSiam: Domain-Aware Siamese Network for Visual Object Tracking

Abstract

Visual object tracking is a fundamental task in the field of computer vision. Recently, Siamese trackers have achieved state-of-theart performance on recent benchmarks. However, Siamese trackers do not fully utilize semantic and objectness information from pre-trained networks that have been trained on the image classification task. Furthermore, the pre-trained Siamese architecture is sparsely activated by the category label which leads to unnecessary calculations and overfitting. In this paper, we propose to learn a Domain-Aware, that is fully utilizing semantic and objectness information while producing a class-agnostic using a ridge regression network. Moreover, to reduce the sparsity problem, we solve the ridge regression problem with a differentiable weighteddynamic loss function. Our tracker, dubbed DomainSiam, improves the feature learning in the training phase and generalization capability to other domains. Extensive experiments are performed on five tracking benchmarks including OTB2013 and OTB2015 for a validation set; as well as the VOT2017, VOT2018, LaSOT, TrackingNet, and GOT10k for a testing set. DomainSiam achieves a state-of-the-art performance on these benchmarks while running at 53 FPS.

SiamMan: Siamese Motion-aware Network for Visual Tracking

“…Our SiamMan method is implemented using the Pytorch tracking platform PySOT 1 . Several experiments are conducted on five challenging datasets, i.e., VOT2016 [20], VOT2018 [21], OTB100 [39], UAV123 [29] and LTB35 [28], to demonstrate the effectiveness of the proposed method.…”

“…To take fully use of semantic information, He et al [16] construct a twofold Siamese networks, which is composed of a semantic branch and an appearance branch, and each of them is a similarity-learning Siamese network. Abdelpakey and Shehat [1] use semantic and objectness information and produce a class-agnostic using a ridge regression network for object tracking.…”

“…Tracker Accuracy Robustness EAO C-COT [11] 0.539 0.238 0.331 SiamRPN [25] 0.560 1.080 0.344 FCAF [15] 0.581 1.020 0.356 C-RPN [13] 0.594 0.950 0.363 SiamRPN+ [41] 0.580 0.240 0.370 ECO [9] 0.550 0.200 0.375 ASRCF [8] 0.563 0.187 0.391 DaSiamRPN [43] 0 [2] 0.503 0.585 0.188 DSiam [14] 0.215 0.646 0.196 SiamRPN [25] 0.490 0.460 0.244 ECO [9] 0.484 0.276 0.280 SA Siam R [16] 0.566 0.258 0.337 DeepSTRCF [26] 0.523 0.215 0.345 DRT [34] 0.519 0.201 0.356 RCO [21] 0.507 0.155 0.376 UPDT [4] 0.536 0.184 0.378 DaSiamRPN [43] 0.586 0.276 0.383 MFT [21] 0.505 0.140 0.385 LADCF [40] 0.503 0.159 0.389 DomainSiam [1] 0.593 0.221 0.396 PTS [37] 0.612 0.220 0.397 ATOM [10] 0 for the non-rigid targets. Compared to SiamRPN++ [24], our SiamMan method produces 0.049 and 0.005 higher EAO and robustness scores, indicating that the localization branch can significantly decrease the tracking failure.…”

“…Modern visual tracking algorithms can be roughly divided into two categories: (1) the correlation filter based approaches [9,26,34,40], and (2) the deep convolution network based approaches [35,24,38,10]. The correlation filter (CF) based approach trains a regressor for tracking using circular correlation via Fast Fourier Transform (FFT).…”