Yan Yan

try to use absolute path for model dir, i will attempt to fix relative path problem later.

"https://github.com/nutonomy/second.pytorch" is pointpillars developed by nutonomy, their code is based on SECOND v1.0. "https://github.com/traveller59/second.pytorch" is SECOND developed by me. the pointpillars in this repo is an **unofficial** implementation, there may...

The all.fhd.config is a multi-class config, but people need small receptive field rather than car, you may need to change architecture for people detection.

[email protected], 0.70, 0.70 means evaluate car performance in easy, moderate, hard and use 0.7 (easy), 0.7 (mod), 0.7 (hard) as overlap threshold. bbox means calculate overlap (intersection of union area)...

the newest code already have pointpillars support. please use ```python setup.py bdist_wheel``` and ```cd ./dist && pip install xxx.whl``` to install spconv.

No, the dependence library ```spconv``` can't be built on windows for now. I just test the spconv build on windows but failed on ```torch.ops.load_library("spconv.dll")```. I'm not familiar with windows build....

I have success build and import the spconv in windows. I will add windows support for spconv soon.

I have tested train/eval in windows 10/cuda 10. The only requirement of windows support is spconv which is solved in newest commit. The major problem of windows is speed, the...

evaluation in car.lite.config in 1060 notebook: Windows 10: ``` generate label finished(27.22/s). start eval: avg example to torch time: 1.558 ms avg prep time: 5.970 ms avg voxel_feature_extractor time =...

I have added some [code](https://github.com/traveller59/second.pytorch/blob/master/second/pytorch/train.py#L588) to ```evaluate``` function in train.py to support cpu inference. very simple. The problem is: 1. inference can run on cpu but eval code only support...