UAST: Uncertainty-Aware Siamese Tracking
Authors: Dawei Zhang, Yanwei Fu, Zhonglong Zheng
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on several challenging tracking benchmarks demonstrate the effectiveness of UAST and its superiority over other Siamese trackers. |
| Researcher Affiliation | Academia | 1College of Mathematics and Computer Science, Zhejiang Normal University, Jinhua, China 2School of Data Science, Fudan University, Shanghai, China 3Key Laboratory of Intelligent Education Technology and Application of Zhejiang Province, Zhejiang Normal University, Jinhua, China. |
| Pseudocode | Yes | Algorithm 1 shows the procedure in details. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing its source code or a link to a code repository for the implemented methodology. |
| Open Datasets | Yes | The training image pairs are sampled by Image Net VID and DET (Russakovsky et al., 2015), COCO (Lin et al., 2014), Youtube-BB (Real et al., 2017), GOT-10K and La SOT (Fan et al., 2019). |
| Dataset Splits | Yes | GOT-10k (Huang et al., 2019) is a large-scale generic object tracking benchmark with 10000 video sequences, which includes 180 videos for testing. Note that it is zero-classoverlap between the train subset and test subset. Following the official protocol, we train UAST only with its training set, and evaluate it with 14 state-of-the-art tracking methods on the test set. |
| Hardware Specification | Yes | Our experiments are conducted on a server with Intel Xeon (R) Gold 5118 CPU, and a Tesla V-100 16 GB GPU. |
| Software Dependencies | Yes | UAST with a speed of 65 fps is implemented by Py Torch 1.1. |
| Experiment Setup | Yes | Template image is 127 127 pixels, while search region is 255 255 pixels. We totally train the network using synchronized stochastic gradient descent (SGD) with a batch size of 128 on 4 GPUs for 20 epochs, and employ warm-up in the first 5 epochs, and a learning rate exponentially decayed from 5e-3 to 1e-6 in the last 15 epochs. We freeze the backbone in the first 10 epochs, and fine-tune it in the remaining epochs. The weight decay and momentum are set as 1e-5 and 0.9, respectively. |