Singular Value Fine-tuning: Few-shot Segmentation requires Few-parameters Fine-tuning

Authors: Yanpeng Sun, Qiang Chen, Xiangyu He, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Jian Cheng, Zechao Li, Jingdong Wang

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our Singular Value Fine-tuning (SVF) approach on various few-shot segmentation methods with different backbones. We achieve state-of-the-art results on both Pascal-5i and COCO-20i across 1-shot and 5-shot settings.
Researcher Affiliation Collaboration Yanpeng Sun1 , Qiang Chen2 , Xiangyu He3 , Jian Wang2, Haocheng Feng2 Junyu Han2, Errui Ding2, Jian Cheng3, Zechao Li1 , Jingdong Wang2 1School of Computer Science and Engineering, Nanjing University of Science and Technology 2Baidu VIS 3NLPR, Institute of Automation, Chinese Academy of Sciences
Pseudocode Yes Algorithm 1 Pseudocode of SVF in Python style
Open Source Code Yes The source code and models will be available at https://github.com/syp2ysy/SVF.
Open Datasets Yes Experiments are conducted on Pascal-5i[29] and COCO-20i [24]. Following the previous work [23, 35, 29], we separate all classes in both datasets into 4 folds.
Dataset Splits Yes In this task, datasets are split into the training set (Dtrain) with base classes (Ctrain) and the testing set (Dtest) with novel classes (Ctest)...
Hardware Specification Yes All model runs on four NVIDIA A100 GPUs.
Software Dependencies No The paper mentions using an "SGD optimizer" but does not specify version numbers for any software libraries, frameworks, or programming languages used.
Experiment Setup Yes We use SGD optimizer with cosine Learning rate decay [21], the learning rate 0.015 and the random seed 321 when fine-tuning backbone. ... All models are trained 200 epochs on Pascal-5i with batch size 8 and trained 50 epochs on COCO-20i with batch size 8. Image is resized to 473 473 on Pascal-5i and 641 641 on COCO-20i.