andyyuan96 / yzlfusion

YZLFusion

Method

plain centerpoint-voxel(without better backbone like focalsparse conv,without better neck like SA-FPN,without better head like transfusionhead) + swin-tiny + fusion strategy.

News

• (2022.9.4) rank 1st and 1st in the term of NDS and mAP on the nuScenes leaderboard among all methods that don't use TTA and Ensemble.
• (2022.9.21) rank 1st and 2nd in the term of NDS and mAP on the nuScenes learderboard among all methods that use single model and TTA.
• (2022.10.6) rank 1st and 3rd in the term of NDS and mAP on the nuScenes learderboard among all methods that use Ensemble. learderboard link

Nuscene Single Model Performance

Valset

Image	Lidar	mAP	NDS
ResNet50	Pillar-02pillar	0.6701	0.6988
ResNet50	VoxelNet-0075voxel	0.7037	0.7259
Swin-Tiny	VoxelNet-0075voxel

Testset

Image	Lidar	mAP	NDS
Swin-Tiny	VoxelNet-0075voxel	0.722	0.738

Nuscene Ensemble Model Performance

Image	Lidar	Strategy	mAP	NDS
Swin-Tiny	VoxelNet-0075	double lidar backbone channel, single-model+flip+rot	0.7363	0.7557

yzlfusion's Issues

Question on YZLFusion implementation

Hi, congratulate to YZLFusion getting great performance on nuScenes dataset.
I wonder that do you conduct cross-modal GT-Paste for YZLFusion? Or simply follows the fade strategy similar to TransFusion (so that no need for GT-Paste)? Besides, are you going to release a paper for YZLFusion?
Thanks!

TTA question

Hi, Thanks for your awesome job!
I have some question about tta, 1) you use double flip or only horizontal direction flip? 2) In my configuration, the improvement from flip TTA is around 1 mAP, but nds don't have improvement, is that correct？thanks so much

About model ensemble

Hi,
Thanks for the great work!
I find that YZLFusion improves NDS by ~2.0 when using the model ensemble strategy, which is a large improvement. I also tried to ensemble some models, including centerpoint, transfusion-L, and transfusion-LC. However, the performance is not good as expected (~0.5NDS improvement).
What models do you choose? Do you only use two models for the ensemble (1. larger backbone 2. models with TTA), as mentioned in the Nuscene Ensemble Model Performance Section? Does a single model with a larger backbone achieve better results than an normal size backbone?
Looking forward to your reply!

About