Optics and Precision Engineering, Volume. 33, Issue 4, 653(2025)
A Transformer-based visual tracker via knowledge distillation
Figures & Tables(10)
Fig. 4. Visualization of feature maps sequentially passing through Conv， BN and ReLU for the predicted frame and current frame
Fig. 6. Comparison with advanced trackers for different attributes on LaSOT
Fig. 7. Screenshots of tracking results for different algorithms on sequences bird1 in OTB
Table 1. Details of KTransT and KTransT-T
View tableView in ArticleTable 1. Details of KTransT and KTransT-T
Method Block Params/M FLOPs/G Latency/ms KTransT-T 12 89 66 10 KTransT 6 46 33 5
Table 2. Ablation experiments on LaSOT
View tableView in ArticleTable 2. Ablation experiments on LaSOT
Method AUC P NP Baseline-T 64.3 69.1 74.0 +OD 65.1 70.2 75.0 +Conv_1 64.5 69.3 74.1 +Conv_3 64.7 69.4 74.5 +OD+Conv_3（KTransT-T） 67.3 72.0 76.6 Baseline-S 55.0 55.7 67.4 +RF 56.0 59.3 68.3 +KD 60.0 62.8 69.7 +KDL 62.5 65.1 71.6 +RF+KDL（KTransT） 64.1 67.9 73.3
Table 3. Comparison of different algorithms on GOT-10k， LaSOT， OTB100， UAV123 and TrackingNet
View tableView in ArticleTable 3. Comparison of different algorithms on GOT-10k， LaSOT， OTB100， UAV123 and TrackingNet
Method GOT-10k LaSOT OTB100 TrackingNet UAV123 GPU Speed AO AUC P AUC P AUC P AUC P SiamFC［ 30 ］35.5 39.0 11.8 33.6 33.9 57.8 76.5 57.1 53.3 52.3 73.1 330 SiamBAN［ 31 ］13.6 8.7 1.4 51.4 52.1 64.0 85.9 - - 55.6 73.3 138 SiamGAT［ 32 ］62.7 74.3 48.8 53.9 53.0 63.1 82.3 75.3 69.8 61.4 81.3 196 ATOM［ 33 ］55.6 63.4 40.2 49.3 48.2 60.3 80.9 70.3 64.8 58.9 79.6 54 DiMP50［ 34 ］61.1 71.7 49.2 56.5 56.9 64.0 85.5 74.0 68.7 59.6 78.6 57 TansT［ 11 ］72.3 82.4 68.2 64.2 68.2 67.2 87.7 81.4 80.3 61.9 79.8 84 E.T.Track［ 35 ］56.8 64.9 42.8 59.0 60.4 67.0 87.3 75.0 70.6 62.7 80.9 64 LightTrack-Mobile［ 36 ］58.2 67.1 44.2 53.8 53.7 66.1 86.0 72.5 69.5 46.9 60.9 71 MixFormerV2-S［ 37 ］61.3 70.5 51.8 60.6 60.4 63.7 84.4 75.8 70.4 65.1 86.6 240 HCAT［ 38 ］65.3 76.8 57 59.1 60.7 68.1 - 76.6 72.9 63.6 - 113 HiT-Tity［ 39 ］52.6 59.3 42.7 54.8 52.9 54.3 70.7 74.6 68.8 58.7 76.6 73 SeqTrack［ 15 ］67.3 76.6 60.7 64.3 69.1 66.2 88.4 79.3 76.6 65.3 85.9 87 KTransT -T 72.4 82.8 68.9 67.3 72.0 67.9 91.0 79.8 86.4 66.2 85.7 91 KTransT 69.8 79.4 64.2 64.1 67.9 64.3 86.0 77.3 80.1 61.8 81.3 158
Tools
Get Citation
Copy Citation Text
Na LI, Mengqiao LIU, Jinting PAN, Kai HUANG, Xingxuan JIA. A Transformer-based visual tracker via knowledge distillation[J]. Optics and Precision Engineering, 2025, 33(4): 653
Paper Information
Category:
Received: Sep. 24, 2024
Accepted: --
Published Online: May. 20, 2025
The Author Email: Na LI (lina114@xupt.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20