Global Dilated Attention and Target Focusing Network for Robust Tracking
Authors: Yun Liang, Qiaoqiao Li, Fumian Long
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on challenging benchmarks (La SOT, Tracking Net, GOT-10k, OTB100) demonstrate that the Gda TFT outperforms many stateof-the-art trackers and achieves leading performance. |
| Researcher Affiliation | Academia | Yun Liang*, Qiaoqiao Li, Fumian Long Guangzhou Key Laboratory of Intelligent Agriculture, College of Mathematics and Informatics, South China Agricultural University yliang@scau.edu.cn, qiaoqiaoli23@163.com, lfm_sunny@stu.scau.edu.cn |
| Pseudocode | No | The paper includes architectural diagrams (e.g., Fig. 3, 4, 6) and descriptive text for its components (GDA, TFN) but does not present any formal pseudocode or algorithm blocks. |
| Open Source Code | No | Code will be available. |
| Open Datasets | Yes | We train Gda TFN with the data from COCO (Lin et al.), GOT-10k (Huang, Zhao, and Huang), Image Net DET/VID (Russakovsky et al.), Tracking Net (Muller et al.) and La SOT (Fan et al.). |
| Dataset Splits | No | The paper mentions using multiple datasets for training (COCO, GOT-10k, Image Net DET/VID, Tracking Net, La SOT) and then evaluating on benchmarks (GOT-10K, OTB100, Tracking Net, La SOT). However, it does not specify explicit train/validation/test splits with percentages or counts for any single dataset, nor does it mention a dedicated validation set or split for hyperparameter tuning separate from the test benchmarks. |
| Hardware Specification | Yes | The proposed Gda TFT is implemented in Python on 4 RTX-2080Ti. |
| Software Dependencies | No | The paper mentions 'Python' as the implementation language but does not specify its version or the versions of any other key software dependencies (e.g., libraries, frameworks, solvers) used in the experiments. |
| Experiment Setup | Yes | During the training, the batchsize is set to 96 and totally 20 epochs are performed by using stochastic gradient descent. The initial learning rate is 1e-6, which increases linearly to 8e-2 within an epoch, and then decreases to 1e-6 for the rest 19 epochs. We employ three loss of focal loss (Lin et al.), binary cross entropy (De Boer et al.) and Io U loss (Yu et al.) to train the model. We combine the three losses with linear process by the ratio 1:1:2. |