Continual Semantic Segmentation Leveraging Image-level Labels and Rehearsal
Authors: Mathieu Pagé Fortin, Brahim Chaib-draa
IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on Pascal-VOC by varying the proportion of fullyand weakly-supervised data in various setups and show that our contributions consistently improve the m Io U on both past and novel classes. |
| Researcher Affiliation | Academia | Mathieu Pag e Fortin , Brahim Chaib-draa Laval University, Qu ebec, Canada mathieu.page-fortin.1@ulaval.ca brahim.chaib-draa@ift.ulaval.ca |
| Pseudocode | No | The paper describes the approach and steps in the main text and through figures, but it does not provide any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain an explicit statement about the release of source code or a link to a code repository. |
| Open Datasets | Yes | We train and evaluate our model in various continual learning scenarios built from the commonly used PASCAL-VOC 2012 dataset [Everingham et al., 2015]. |
| Dataset Splits | Yes | This dataset contains a train split of 10,582 images and a val split of 1,449 images used for testing. ... the hyper-parameters are searched by keeping 20% of the training set for validation. |
| Hardware Specification | Yes | We train our models with SGD with momentum on 4 Nvidia A100 GPUs with a total batch size of 24 for 30 epochs for each step. |
| Software Dependencies | No | The paper mentions several components like Deeplab-v3, ResNet-101, and ImageNet, but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | We train our models with SGD with momentum on 4 Nvidia A100 GPUs with a total batch size of 24 for 30 epochs for each step. An initial learning rate of 0.01 and 0.001 are used for the first and subsequent steps, respectively, with a polynomial decay of power 0.9. ...Additionally, with weaklysupervised data a threshold of 0.75 is used to only keep the most confident predictions as pseudo-labels. The weights for distillation losses are 100 in the 19-1 and 15-1 scenarios, and 10 for the 15-5 scenarios. |