search

traveller59 / second.pytorch

SECOND for KITTI/NuScenes object detection
MIT License
1.72k stars 722 forks source link

train and inference using apollo dataset #270

Open 940023303 opened 5 years ago

940023303 commented 5 years ago

@traveller59 I want to train the model using Apollo dataset and inference my own dataset with trained model. now I have modified some code and the process of training seems correct, but I don't know how to modify the evalution function in dataset class. Apollo dataset and my own dataset just have pointcloud and label include type and dimension and location and heading. I inference the kitti dataset using the model trained with apolo dataset. This is the result. Evaluation official Car AP(Average Precision)@0.70, 0.70, 0.70: bbox AP:33.25, 24.56, 21.84 bev AP:37.20, 26.45, 21.42 3d AP:9.70, 9.09, 9.09 aos AP:16.76, 12.39, 11.04 Car AP(Average Precision)@0.70, 0.50, 0.50: bbox AP:33.25, 24.56, 21.84 bev AP:55.23, 39.50, 33.70 3d AP:50.94, 36.79, 30.99 aos AP:16.76, 12.39, 11.04

Evaluation coco Car coco AP@0.50:0.05:0.95: bbox AP:29.07, 21.98, 19.82 bev AP:29.14, 21.61, 18.36 3d AP:16.31, 12.79, 11.27 aos AP:14.63, 11.09, 10.02

this is the url of downloading the apollo dataset. http://apolloscape.auto/tracking.html

Can you help me or give me some advise?

AddASecond commented 5 years ago

Same question. Have you solved it yet? @940023303

940023303 commented 5 years ago

@AddASecond yes，i write a new function to evaluate the detection performance.

AddASecond commented 5 years ago

@AddASecond yes，i write a new function to evaluate the detection performance.

How much time did you train your model? For Apollo, I only got 14AP in Car, using one P5000(only can batchsize 1) for 3 days.

940023303 commented 5 years ago

How much time did you train your model? For Apollo, I only got 14AP in Car, using one P5000(only can batchsize 1) for 3 days.

@AddASecond mybe six hours with 1080Ti. the average iou is near 50%.

lonlonago commented 5 years ago

@940023303 what‘s your ap like? could you share the evau code?

ryontang commented 3 years ago

@AddASecond Hi, Dose it have some differences between training by kitti data and apollo data.

CaptainCSD commented 3 years ago

@AddASecond yes，i write a new function to evaluate the detection performance.

How much time did you train your model? For Apollo, I only got 14AP in Car, using one P5000(only can batchsize 1) for 3 days.

wtf man, 3days? I am using GTX 3080 for training with setting batch size of 6, any idea of how long it will take to finish the training ?