409caf8548
## Motivation We are from NVIDIA and we have developed a simplified and inference-efficient transformer for dense prediction tasks. The method is based on SegFormer with hardware-friendly design choices, resulting in better accuracy and over 2x reduction in inference speed as compared to the baseline. We believe this model would be of particular interests to those who want to deploy an efficient vision transformer for production, and it is easily adaptable to other tasks. Therefore, we would like to contribute our method to mmsegmentation in order to benefit a larger audience. The paper was accepted to [Transformer for Vision workshop](https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fsites.google.com%2Fview%2Ft4v-cvpr22%2Fpapers%3Fauthuser%3D0&data=05%7C01%7Cboyinz%40nvidia.com%7Cbf078d69821449d1f4c908dab5e8c7da%7C43083d15727340c1b7db39efd9ccc17a%7C0%7C0%7C638022308636438546%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=XtSgPQrbVgHxt5L9XkXF%2BGWvc95haB3kKPcHnsVIF3M%3D&reserved=0) at CVPR 2022, here below are some resource links: Paper [https://arxiv.org/pdf/2204.13791.pdf](https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Farxiv.org%2Fpdf%2F2204.13791.pdf&data=05%7C01%7Cboyinz%40nvidia.com%7Cbf078d69821449d1f4c908dab5e8c7da%7C43083d15727340c1b7db39efd9ccc17a%7C0%7C0%7C638022308636438546%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=X%2FCVoa6PFA09EHfClES36QOa5NvbZu%2F6IDfBVwiYywU%3D&reserved=0) (Table 3 shows the semseg results) Code [https://github.com/NVIDIA/DL4AGX/tree/master/DEST](https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FNVIDIA%2FDL4AGX%2Ftree%2Fmaster%2FDEST&data=05%7C01%7Cboyinz%40nvidia.com%7Cbf078d69821449d1f4c908dab5e8c7da%7C43083d15727340c1b7db39efd9ccc17a%7C0%7C0%7C638022308636438546%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=9DLQZpEq1cN75%2FDf%2FniUOOUFS1ABX8FEUH02O6isGVQ%3D&reserved=0) A webinar on its application [https://www.nvidia.com/en-us/on-demand/session/other2022-drivetraining/](https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.nvidia.com%2Fen-us%2Fon-demand%2Fsession%2Fother2022-drivetraining%2F&data=05%7C01%7Cboyinz%40nvidia.com%7Cbf078d69821449d1f4c908dab5e8c7da%7C43083d15727340c1b7db39efd9ccc17a%7C0%7C0%7C638022308636438546%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=8jrBC%2Bp3jGxiaW4vtSfhh6GozC3tRqGNjNoALM%2FOYxs%3D&reserved=0) ## Modification Add backbone(smit.py) and head(dest_head.py) of DEST ## BC-breaking (Optional) N/A ## Use cases (Optional) N/A --------- Co-authored-by: MeowZheng <meowzheng@outlook.com>
Projects
Implementing new models and features into OpenMMLab's algorithm libraries could be troublesome due to the rigorous requirements on code quality, which could hinder the fast iteration of SOTA models and might discourage our members from sharing their latest outcomes here.
And that's why we have this Projects/ folder now, where some experimental features, frameworks and models are placed, only needed to satisfy the minimum requirement on the code quality, and can be used as standalone libraries. Users are welcome to use them if they use MMSegmentation from source.
Everyone is welcome to post their implementation of any great ideas in this folder! If you wish to start your own project, please go through the example project for the best practice.
Note: The core maintainers of MMSegmentation only ensure the results are reproducible and the code quality meets its claim at the time each project was submitted, but they may not be responsible for future maintenance. The original authors take responsibility for maintaining their own projects.