SSUL: Semantic Segmentation with Unknown Label for Exemplar-based Class-Incremental Learning
Authors: Sungmin Cha, beomyoung kim, YoungJoon Yoo, Taesup Moon
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The extensively conducted experiments show the effectiveness of our method, achieving significantly better performance than the recent state-of-the-art baselines on the standard benchmark datasets. Furthermore, we justify our contributions with thorough ablation analyses and discuss different natures of the CISS problem compared to the traditional class-incremental learning targeting classification. |
| Researcher Affiliation | Collaboration | 1 Department of Electrical and Computer Engineering, Seoul National University 2 NAVER AI Lab, 3 Face, NAVER Clova |
| Pseudocode | Yes | Algorithm 1: Pseudo code of SSUL-M |
| Open Source Code | Yes | The official code is available at https://github.com/clovaai/SSUL. |
| Open Datasets | Yes | We followed the experimental setting of [3] and evaluated our method using Pascal-VOC 2012 [9] and ADE20K [32] datasets. [9] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303 338, 2010. [32] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633 641, 2017. |
| Dataset Splits | Yes | We tune the hyperparameters using 20% of the training set as a validation set and report the final results on the standard test set. |
| Hardware Specification | Yes | The experiments were implemented in Py Torch [26] 1.7 with CUDA 10.1 and Cu DNN 7 using two NVIDIA V100 GPUs, and all experiments were conducted with NSML [15] framework. |
| Software Dependencies | Yes | The experiments were implemented in Py Torch [26] 1.7 with CUDA 10.1 and Cu DNN 7 using two NVIDIA V100 GPUs, and all experiments were conducted with NSML [15] framework. |
| Experiment Setup | Yes | For all experiments, following other works [7, 3], we use a Deeplabv3 segmentation network [5] with a Res Net-101 [10] backbone pre-trained on Image Net [6]. We optimize the network using SGD with an initial learning rate of 10 2 and a momentum value of 0.9 for all CISS steps. Also, we set the learning rate schedule, data augmentation, and output stride of 16 following [5] for all experiments. For each incremental state t, we train the network for 50 epochs for Pascal VOC with a batch-size of 32 and 60 epochs for ADE20K with a batch-size of 24. We tune the hyperparameters using 20% of the training set as a validation set and report the final results on the standard test set. For the exemplar memory, we utilized memory with a fixed size of |M| = 100 for Pascal VOC and |M| = 300 for ADE20K. |