Lightweight Frequency Masker for Cross-Domain Few-Shot Semantic Segmentation
Authors: Jintao Tong, Yixiong Zou, Yuhua Li, Ruixuan Li
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on four target datasets show that our work, through a simple and effective design, significantly outperforms the state-of-the-art CD-FSS method. |
| Researcher Affiliation | Academia | Jintao Tong Yixiong Zou Yuhua Li Ruixuan Li School of Computer Science and Technology, Huazhong University of Science and Technology {jintaotong, yixiongz, idcliyuhua, rxli}@hust.edu.cn |
| Pseudocode | No | The paper uses mathematical formulations and flowcharts (Figure 4) to describe its modules and operations, but does not include any explicitly labeled "Pseudocode" or "Algorithm" blocks. |
| Open Source Code | Yes | Corresponding author. Code is available at https://github.com/Tung Chintao/APM. |
| Open Datasets | Yes | For training, our source domain is the PASCAL-5i dataset [35], an extended version of PASCAL VOC 2012 [13] enhanced with additional annotations from the SDS dataset. For evaluation, our target domains include FSS-1000 [26], Deepglobe [10], ISIC2018 [9, 38], and the Chest X-ray datasets [6, 20]. |
| Dataset Splits | No | In this work, we adopt the episodic training manner. Specifically, both the training set sampled from Ds and the testing set sampled from Dt are composed of several episodes, each episode is constructed of K support samples S = {Ii s, M i s}K i=1 and a query Q = {Iq, Mq} (I is the image and M is the label). |
| Hardware Specification | No | The implementation details in Section 4.2 do not specify any hardware (GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using the Adam [21] optimizer but does not specify software dependencies like specific deep learning frameworks (e.g., PyTorch, TensorFlow) or their version numbers. |
| Experiment Setup | Yes | The spatial sizes of both support and query images are set to 400 400. The model is trained using the Adam [21] optimizer with a learning rate of 1e-3. ... learning rates set at 0.1 for Chest X-ray, 0.01 for FSS-1000 and ISIC, and 1e-5 for Deepglobe. Each task undergoes a total of 60 iterations. |