search

charlesq34 / frustum-pointnets

Frustum PointNets for 3D Object Detection from RGB-D Data
Apache License 2.0
1.57k stars 538 forks source link

Cannot reproduce the reported result on kitti val dataset(car only) #88

Open lhyfst opened 5 years ago

lhyfst commented 5 years ago

I only generate car only data when I prepare the data. Then I use the car only data to train and test the v2 model(using default hyper-parameters). However, I cannot reproduce the result reported in your paper. Did you modify the training code or model code when you trained on car only data? Or is there a set of special hyper-parameters which suits for car only data? Thanks!

Here is the report on car only data. image

Here is the result I got using default hyper-parameters. car_detection_ground AP: 84.5136 74.8162 65.6415 car_detection_3d AP: 69.0851 55.0259 47.6502

malicd commented 4 years ago

Did you change any of the hyper parameters yourself (i.e. batch size)? I trained the model for 190 epochs and got following results:

car_detection_ground AP: 87.719498 82.619095 75.542664
car_detection_3d AP: 82.550797 69.479378 62.767441

which is very similar to what they report in the paper.

kobe-wei commented 3 years ago

Did you change any of the hyper parameters yourself (i.e. batch size)? I trained the model for 190 epochs and got following results:

car_detection_ground AP: 87.719498 82.619095 75.542664
car_detection_3d AP: 82.550797 69.479378 62.767441

which is very similar to what they report in the paper.

hi, why my result on kitti val dataset (car only) using author's model is 0, 0, 0? please tell me where is wrong. Update:I have solved this problem caused by the circulation in inference.And the car_detection_ground AP: 88.095261 82.646202 74.777641, car_detection_3d AP: 84.562439 71.432144 63.435196.