Is there any detection result available ? about second.pytorch

[traveller59]

What kind of detection results do you want? KITTI label files of valid/test set? Images/pointcloud images with detection?
you can use python ./pytorch/train.py evaluate ... --pickle_result=False to generate label files, use kitti viewer to visualize detection results in GUI.

[Oofs]

I had tried ./pytorch/train.py evaluate to generate detection result, but unfortunately my GPU had ran out of memory...(even batch_size=1 is too large)
I also tried to use kitti viewer to inference one pointcloud using pretrained model. I click load info and then click load and inferenct network, but it just quit and give me following information :

load 14357 Car database infos
load 2207 Pedestrian database infos
load 734 Cyclist database infos
load 1297 Van database infos
load 56 Person_sitting database infos
load 488 Truck database infos
load 224 Tram database infos
load 337 Misc database infos
[ 11 400 352]
Restoring parameters from /home/tianning/oofs/second.pytorch/pretrained_models/car/voxelnet-204270.tckpt
Traceback (most recent call last):
File "kittiviewer/viewer.py", line 1319, in on_LoadInferenceVxNetPressed
self.on_InferenceVxNetPressed()
File "kittiviewer/viewer.py", line 1308, in on_InferenceVxNetPressed
self.kitti_info, self.points)
File "/home/tianning/oofs/second.pytorch/second/core/inference.py", line 26, in get_inference_input_dict
rect = info['calib/R0_rect']
TypeError: 'NoneType' object is not subscriptable

So i guess maybe i lack some files ? Do you have any idea about this?

[traveller59]

maybe you need to click plot first. If you can't run evaluate, you may can't run inference in GUI too.

[traveller59]

I will upload a smaller model in one day, maybe you can use it to evaluate/visualize.

[Oofs]

Get it worked with GTX 1080:

I'm planning to use this net in ROS to accomplish real-time detection with vlp-16. Have you ever tried VLP-16 data with your net ?

[traveller59]

I have used this network to detect trucks in a private two-16-lidar dataset, I don't know if it possible to detect cars in complex environment with only one 16-line lidar.

[kwea123]

@Oofs Hi, can you kindly record some video showing it in action? like this.
I wanted to test the model, but my PC works very slow. Thank you for your kindness!

[kxhit]

Hi! @traveller59 I have successfully trained and evaluated the model. Thanks for your great work!
Following your instructions, I meet some problems during the visualization.

1. The Groundtruth Boxes are not drawed.
2. Could we visualize multiple categories at once? There only draws car detection results.
3. I notice that there only 3 configs(car,car.tiny,people) in configs folder. Could we visualize other categories? And how many categories have you trained?

Thanks! Waiting for your reply!

[kwea123]

@traveller59 How does this network perform on 16 lines data?

[minghaohsu410168]

@Oofs Excuse me, have you use this net in ROS to accomplish real-time detection with vlp-16? Would you give me some advise, thank you so much.