search

stjuliet / CenterLoc3D

The offical implementation (pytorch) of CenterLoc3D: Monocular 3D Vehicle Localization Network for Roadside Surveillance Cameras
https://arxiv.org/abs/2203.14550
MIT License
12 stars 2 forks source link

Where is the pre-train weight? #1

Open TienDungHUST opened 2 years ago

TienDungHUST commented 2 years ago

Hi, Thank you for your research. I'm following your instruction to run the predict.single_predict.py. Howerver, I can't find the weight file '../0all-module-logs/resnet50-Epoch99-ciou-Total_train_Loss1.5854-Val_Loss2.2824.pth'. Can you share this weight, please?

TienDungHUST commented 2 years ago

I re-trained myself using your dataset and get the weight resnet50-Epoch96-ciou-Total_train_Loss1.8821-Val_Loss2.3004.pth. It looks like your result, however, it seem that the model becomes overfitted. The 3D bbox detections are wrong on my own test image, even when I test with cropped image from SVLD-TEST

wxlwhu commented 1 year ago

resnet50-Epoch99-ciou-Total_train_Loss1.5854-Val_Loss2.2824.pth

Hello,did you finally find the wight files?

TienDungHUST commented 1 year ago

resnet50-Epoch99-ciou-Total_train_Loss1.5854-Val_Loss2.2824.pth

Hello,did you finally find the wight files?

Hi, I did not find it. But you can train it yourself, it takes short time to get convergence. Note that, the model will be overfitting with the road scenes which you have in the training set.