Class-Incremental Instance Segmentation via Multi-Teacher Networks

Authors: Yanan Gu, Cheng Deng, Kun Wei1478-1486

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results on PASCAL 2012 SBD and COCO datasets show the effectiveness of the proposed method. And we evaluate variants of our method with ablation studies to verify the effectiveness of each component in our model.
Researcher Affiliation Academia Yanan Gu, Cheng Deng , and Kun Wei School of Electronic Engineering, Xidian University, Xi an 710071, China
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not contain a statement about open-sourcing the code or a link to a code repository.
Open Datasets Yes We evaluate our method on the PASCAL 2012 SBD (Hariharan et al. 2011) dataset and Microsoft COCO (Lin et al. 2014) dataset.
Dataset Splits Yes On the other hand, COCO has 80k images in the training set and 40k images in the validation set.
Hardware Specification No The paper does not provide specific details about the hardware used, such as GPU or CPU models, or memory specifications.
Software Dependencies No The paper mentions several frameworks and models (e.g., YOLACT, ResNet-50) but does not provide specific version numbers for software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes We use SGD (Bottou 2012) to train all the models in the experiments. The backbone in YOLACT is Res Net-50 (He et al. 2016) and the base image size is 550 550. For the first task, we set the learning rate to 1e-4, and for the next tasks, we set the learning rate to 5e-5. For PASCAL 2012 SBD dataset, we train the model for 120k iterations, and the learning rate will decay at iterations 60k and 100k using a weight decay of 5e-4. For COCO dataset, the model is trained for 400k iterations, and the learning rate decays at iterations 140k, 300k, and 350k. The ratio of the three parameters λ1,λ2, and λ3 determine the importance of three different distillation losses. And the ratio of the three parameters β1,β2, and β3 decide the balance between the learning of old and new classes.