search

hailanyi / TED

Transformation-Equivariant 3D Object Detection for Autonomous Driving
https://arxiv.org/abs/2211.11962
Apache License 2.0
136 stars 32 forks source link

The inference time of TED-M #34

Closed Feidashen1 closed 9 months ago

Feidashen1 commented 9 months ago

Hello author，I had a problem replicating your code. In the paper, the inference speed of TED-M is superior to SFD, but in my reproduction experiment, TEM-M performs worse than SFD, even worse (SFD is 0.13s per frame, TEM-M is 2.01s per frame). What is the possible reason? thanks!

hailanyi commented 9 months ago

To ensure better and stable performance, the released code is somewhat different from paper. The released code is based on multi-stage refinement, which require three RCNN detection heads. While the paper uses only single RCNN head for refinement. So the speed of this code is lower than paper but the accuracy is better than paper.

Feidashen1 commented 9 months ago

To ensure better and stable performance, the released code is somewhat different from paper. The released code is based on multi-stage refinement, which require three RCNN detection heads. While the paper uses only single RCNN head for refinement. So the speed of this code is lower than paper but the accuracy is better than paper.

Thank you for your reply! Can you provide a code or modification method for the RCNN detection head？Because I need to compare it with TED-M in my comparative experiment and measure its accuracy and inference time simultaneously. If I use the released code you provided now, it may not be very accurate.