mmdetection3d
mmdetection3d copied to clipboard
open-mmlab
DeepFusion implementation plans?
DeepFusion 3D object detector
I was wondering if there are any plans to include DeepFusion lidar-camera 3D object detector architecture from Google into MMDetection3D. There is a recently released TensorFlow code as part of LingVO that might aid in reproducibility. It would be wonderful to count on that architecture in this outstanding library.
Related resources
- Link to paper: https://arxiv.org/abs/2203.08195
- Link to LingVO implementation: https://github.com/tensorflow/lingvo/blob/master/lingvo/tasks/car/deep_fusion.py
Great suggestion. We will consider adding this feature in the future but currently, we may not have enough resources to support it very soon. You can refer to other state-of-the-art multi-modality methods (such as TransFusion, BEVFusion or UVTR) based on mmdet3d temporarily.
Great suggestion. We will consider adding this feature in the future but currently, we may not have enough resources to support it very soon. You can refer to other state-of-the-art multi-modality methods (such as TransFusion, BEVFusion or UVTR) based on mmdet3d temporarily.
I already implement deep fusion use mmdet3d 1.0, and I only verify the performance on nuscene dataset, If needed, maybe I can contribute the method to mmdet3d, but it need time to modify code to satisfy review, and I'm not the official implementation.
Great! Contributions are definitely welcome. Feel free to create PRs when you are available, and let us know if you need any help.
@AndyYuan96 Nice to know this. I'm also doing this, but I don't have enough experience in mmdetection3d, so progress is slow. Do you have any plan when you will PR or pre-release on your personal Github?
Great suggestion. We will consider adding this feature in the future but currently, we may not have enough resources to support it very soon. You can refer to other state-of-the-art multi-modality methods (such as TransFusion, BEVFusion or UVTR) based on mmdet3d temporarily.
I already implement deep fusion use mmdet3d 1.0, and I only verify the performance on nuscene dataset, If needed, maybe I can contribute the method to mmdet3d, but it need time to modify code to satisfy review, and I'm not the official implementation.
any advance on this front? I'm also interested in seeing & using DeepFusion from MMdetection3D.
could you share your deepfussion modules/changes for us to see/review? thanks in advance
