ReMaX: relaxing for better training on efficient panoptic segmentation
This paper presents a new mechanism to facilitate the training of mask transformers for efficient panoptic segmentation, democratizing its deployment. We observe that due to the high complexity in the training objective of panoptic segmentation, it will inevitably lead to much higher penalization on...
| Main Authors: | , , , , , |
|---|---|
| Format: | Conference item |
| Language: | English |
| Published: |
NeurIPS
2023
|
| _version_ | 1826312984082776064 |
|---|---|
| author | Sun, S Wang, W Yu, Q Howard, A Torr, P Chen, L-C |
| author_facet | Sun, S Wang, W Yu, Q Howard, A Torr, P Chen, L-C |
| author_sort | Sun, S |
| collection | OXFORD |
| description | This paper presents a new mechanism to facilitate the training of mask transformers for efficient panoptic segmentation, democratizing its deployment. We observe that due to the high complexity in the training objective of panoptic segmentation, it will inevitably lead to much higher penalization on false positive. Such unbalanced loss makes the training process of the end-to-end mask-transformer based architectures difficult, especially for efficient models. In this paper, we present ReMaX that adds relaxation to mask predictions and class predictions during the training phase for panoptic segmentation. We demonstrate that via these simple relaxation techniques during training, our model can be consistently improved by a clear margin without any extra computational cost on inference. By combining our method with efficient backbones like MobileNetV3-Small, our method achieves new state-of-the-art results for efficient panoptic segmentation on COCO, ADE20K and Cityscapes. Code and pre-trained checkpoints will be available at https://github.com/google-research/deeplab2. |
| first_indexed | 2024-09-25T04:03:50Z |
| format | Conference item |
| id | oxford-uuid:98ec7151-ea71-4442-b18c-40aa8001aa4b |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-09-25T04:03:50Z |
| publishDate | 2023 |
| publisher | NeurIPS |
| record_format | dspace |
| spelling | oxford-uuid:98ec7151-ea71-4442-b18c-40aa8001aa4b2024-05-16T15:24:23ZReMaX: relaxing for better training on efficient panoptic segmentationConference itemhttp://purl.org/coar/resource_type/c_5794uuid:98ec7151-ea71-4442-b18c-40aa8001aa4bEnglishSymplectic ElementsNeurIPS2023Sun, SWang, WYu, QHoward, ATorr, PChen, L-CThis paper presents a new mechanism to facilitate the training of mask transformers for efficient panoptic segmentation, democratizing its deployment. We observe that due to the high complexity in the training objective of panoptic segmentation, it will inevitably lead to much higher penalization on false positive. Such unbalanced loss makes the training process of the end-to-end mask-transformer based architectures difficult, especially for efficient models. In this paper, we present ReMaX that adds relaxation to mask predictions and class predictions during the training phase for panoptic segmentation. We demonstrate that via these simple relaxation techniques during training, our model can be consistently improved by a clear margin without any extra computational cost on inference. By combining our method with efficient backbones like MobileNetV3-Small, our method achieves new state-of-the-art results for efficient panoptic segmentation on COCO, ADE20K and Cityscapes. Code and pre-trained checkpoints will be available at https://github.com/google-research/deeplab2. |
| spellingShingle | Sun, S Wang, W Yu, Q Howard, A Torr, P Chen, L-C ReMaX: relaxing for better training on efficient panoptic segmentation |
| title | ReMaX: relaxing for better training on efficient panoptic segmentation |
| title_full | ReMaX: relaxing for better training on efficient panoptic segmentation |
| title_fullStr | ReMaX: relaxing for better training on efficient panoptic segmentation |
| title_full_unstemmed | ReMaX: relaxing for better training on efficient panoptic segmentation |
| title_short | ReMaX: relaxing for better training on efficient panoptic segmentation |
| title_sort | remax relaxing for better training on efficient panoptic segmentation |
| work_keys_str_mv | AT suns remaxrelaxingforbettertrainingonefficientpanopticsegmentation AT wangw remaxrelaxingforbettertrainingonefficientpanopticsegmentation AT yuq remaxrelaxingforbettertrainingonefficientpanopticsegmentation AT howarda remaxrelaxingforbettertrainingonefficientpanopticsegmentation AT torrp remaxrelaxingforbettertrainingonefficientpanopticsegmentation AT chenlc remaxrelaxingforbettertrainingonefficientpanopticsegmentation |