horizon-research/efficient-deep-learning-for-point-clouds

efficient-deep-learning-for-point-clouds's Issues

f-pointnet：a big gap between test in my environment and the your screenshot,can you help me find the problem?

ubuntu16.04
python 2.7
tensorflow1.9.0 cuda9.0 cudnn 7.6.5 numpy 1.16.6. I use your model

in my environment Fully Delayed-Aggregation: | your result
Going to eval ground for class: car
save detection_results_v2/plot/car_detection_ground.txt
car_detection_ground AP: 85.563286 76.262955 72.546509 | 86.406395 81.970383 74.64431
Going to eval ground for class: pedestrian
save detection_results_v2/plot/pedestrian_detection_ground.txt
pedestrian_detection_ground AP: 57.766247 52.338993 45.705994 | 72.302208 66.122513 59.363037
Going to eval ground for class: cyclist
save detection_results_v2/plot/cyclist_detection_ground.txt
cyclist_detection_ground AP: 61.605274 45.203564 42.396843 | 84.097448 64.391205 60.237617
Finished Birdeye eval.

I directly use the training and eval data downloaded long ago when I learn f-pointnet? I maybe make a error ,can you help me find the problem?

Thank you very much?

Low training accuracy with pointnet2

Hello,

I installed your codes with the pointnet2 dataset
Thanks for your kind README, I easily did build the mesorasi code!

However, I cannot do training rightly

python launch.py --download pointnet2
python launch.py --compile pointnet2
python launch.py --train pointnet2

I did the above three commands, and there is no error

This is the result of compile

This is a screenshot of training

During 200 epochs, accuracy did not become higher.
Can you advise me on training?

I think training does not work well.

My setup is

• Ubuntu 18.04
• CUDA 10.2
• cudnn 7.6.5
• tensorflow-gpu 1.14.0
• g++ 4.8

fpointnet has no training and evaluation instructions for limited DA.

refactor the code so that limited-DA doesn't require re-training

DGCNN cls has no training instructions

the question about the paper in IV-A

hello, I was really impressed on the idea of delayed-aggregation, and the experiment seem the module accuracy suffer only a litter effect, but reduce a lot of calculation work, cool!
but I also confuse about the Equ.3, and I can't deducing it. could you help me?

e.g.

φ(φ(5-6)*4)*3) = 0
φ(φ(5)*4)*3) - φ(φ(6)*4)*3) =  - 12

i think maybe
φ(φ(P-Pi)*w1)*w2) = φ { φ(φ(P)*w1)*w2)- φ(φ(Pi)*w1)*w2)}
make more sense.

DGCNN seg has no training instructions

fpointnet has no screenshot for limited DA. the screenshot contains too many meaningless details.

pointnet2 seg has no training instructions.

Runtime: I don't know why the time consumption is basically the same as f-pointnet(baseline) 0.011-0.015 s

imgs