Submit your tracking results on GOT-10k to this website and evaluate the performance. Leaderboard will be updated immediately once the results are evaluated.
Ensembel Siamese Netwroks
We introduce ensemble learning into SiamFC to improve the discriminative ability of the tracker by integrating multiple base trackers, which has strong discriminative ability to identify specific similar target objects into a strong tracker. In detail, we extract the features of the whole dataset, and cluster the target objects according to their distances in the feature space to obtain k clusters. Then, on these clusters, we perform transfer learning on SiamFC to obtain base trackers with specific preferences in discrimination ability. As shown in Fig. 1, in order to better integrate these base trackers with diversity, we design the Cluster Weight module (CW). This module can predict the reliability of each base tracker in the current scene based on the semantic information of the target object and adaptively assign weights to each base tracker. The final response map is obtained by weighted summation of the base tracker response maps, and the scale and position of the target are determined according to the maximum response score of the final response map.
E5-2697v3(ES) GTX 1070
Ensemble Siamese Networks for Object Tracking
|1||ESiamFC||0.381||0.435||0.132||26.36 fps||E5-2697v3(ES) GTX 1070||Python||2021-05-05, 02:38:28||results.zip||reports.json|