Closed shanqiu24 closed 1 year ago
Hello😊
I will share the experimental results for both questions here.
Hello😊
- We have used ImageNet pretrained weights on the DLA-34 backbone as in the original code. I am also curious about the overfitting problem you mentioned, so I am training from scratch. However, my opinion is that there is a risk of overfitting because the KITTI 3D Object Detection dataset is not that large.
- You're right. We trained and tested using only the DLA-34 backbone. I am training the model using pretrained DLA-102 backbone. This experiment will take longer. (This is because DLA-34 has 19 million parameters and DLA-102 has 49 million parameters.)
I will share the experimental results for both questions here.
Got it,I appreciate it~😊
Thank you for waiting.
The experimental results are out, please refer to the table below.
For a fair comparison, the seed of all experiments was fixed to 0.
Please understand that DLA34 (Pretrained) is not a reproduced result. Because the best performance written in the README is the result obtained from a random seed.
AP40@Easy | AP40@Mod. | AP40@Hard | |
---|---|---|---|
Official | 26.33 | 19.03 | 16.00 |
DLA34 (Pretrained) | 24.29 | 17.95 | 15.24 |
DLA34 (Non-Pretrained) | 22.71 | 16.72 | 14.00 |
DLA102 (Pretrained) | 25.33 | 18.23 | 15.41 |
It can be seen that there is a performance gap between DLA34 (Pretrained) and DLA34 (Non-Pretrained).
So, unless you have sophisticated training tricks, it seems good to use pre-trained weights.
Comparing DLA34 (Pretrained) and DLA102 (Pretrained), the performance when using a heavier model is better, but the difference is not large. Therefore, in my opinion, using DLA-34 like other papers currently published would be the most appropriate considering the trade-off between training time and performance.
Thank you for waiting. The experimental results are out, please refer to the table below. For a fair comparison, the seed of all experiments was fixed to 0.
Please understand that DLA34 (Pretrained) is not a reproduced result. Because the performance written in the README is the result obtained from a random seed.
3D Metric on Car Class
AP40@Easy AP40@Mod. AP40@Hard Official 26.33 19.03 16.00 DLA34 (Pretrained) 24.29 17.95 15.24 DLA34 (Non-Pretrained) 22.71 16.72 14.00 DLA102 (Pretrained) 25.33 18.23 15.41
Q1.
It can be seen that there is a performance gap between DLA34 (Pretrained) and DLA34 (Non-Pretrained). So, unless you have sophisticated training tricks, it seems good to use pre-trained weights.
Q2.
Comparing DLA34 (Pretrained) and DLA102 (Pretrained), the performance when using a heavier model is better, but the difference is not large. Therefore, in my opinion, using DLA-34 like other papers currently published would be the most appropriate considering the trade-off between training time and performance.
oh bro ,i got it ,thanks alot~😊
Hey guys,I come again,the pre-trained Backbone must work better than a zero-based one? I mean starting from scratch, the overfitting problem is unavoidable? And the best number of layers you've tried is 34, right? Have you ever tried 100+ DLA?