lordliang / fvnet

FVNet is proposed for 3D object front-view proposal generation and detection from point cloud

FVNet

FVNet is proposed for 3D object front-view proposal generation and detection from point cloud

Previous coarse research paper: FVNet

This just is my graduation project, and I will complete it gradually. I promise. ^_^

fvnet's Issues

Need list_files directory

Could you commit the list_files, such as det_train_car_filtered.txt, label_train_2_car_filtered.txt. Thank you!

Request for much more information to run your code

I've read your paper: FVNet, and I eagerly want to get much more information from your README to run your code and see some results, pls.

Where is the implementing functions of PG-Net?

As a novice, during learning your codes, I have found the implementing functions of PE-Net module, such as get_center_regression_net(), get_3d_box_estimation_net() in front_pointnets_v1.py. But I totally can't find the implementing functions of PG-Net. Could you give me some hints? Thanks a lot!

Some questions after running val.py

1. The results eval box estimation accuracy is always 0
   in log_test_val.py
   Namespace(batch_size=32, gpu=0, model='front_pointnets_v1', model_path='/home/FVNet/est-kittinet/outputs/20191219100152/model_1000.ckpt', num_point=512, output_dir='/home/FVNet/est-kittinet/prediction/val')
   pid: 7425
   eval mean total loss: 7090.168824
   eval mean center loss: 2997.096728
   eval mean stage1 center loss: 2894.534797
   eval mean angle class loss: 1.750437
   eval mean angle res loss: 0.164404
   eval mean size res loss: 4.262741
   eval mean corners loss: 221.648795
   eval box IoU (ground/3D): 0.000003 / 0.000000
   eval box estimation accuracy (IoU=0.5): 0.000000
   eval box estimation accuracy (IoU=0.7): 0.000000

The same in log_train.txt:
**** EPOCH 999 ****
2019-12-20 01:50:44.774636
-- 400 / 429 --
mean total loss: 4238.088100
mean center loss: 1508.271276
mean stage1 center loss: 2049.937858
mean angle class loss: 1.753821
mean angle res loss: 0.164341
mean size res loss: 4.236634
mean corners loss: 118.021132
box IoU (ground/3D): 0.000000 / 0.000000
box estimation accuracy (IoU=0.5): 0.000000
box estimation accuracy (IoU=0.7): 0.000000
**** EPOCH 1000 ****
2019-12-20 01:51:33.322000
-- 400 / 429 --
mean total loss: 4242.318448
mean center loss: 1517.564768
mean stage1 center loss: 2046.396386
mean angle class loss: 1.752404
mean angle res loss: 0.164301
mean size res loss: 4.221072
mean corners loss: 117.779481
box IoU (ground/3D): 0.000000 / 0.000000
box estimation accuracy (IoU=0.5): 0.000000
box estimation accuracy (IoU=0.7): 0.000000
2019-12-20 01:52:21.722286
---- EPOCH 199 EVALUATION ----
eval mean total loss: 7092.537344
eval mean center loss: 2995.654283
eval mean stage1 center loss: 2896.207265
eval mean angle class loss: 1.749893
eval mean angle res loss: 0.164299
eval mean size res loss: 4.265863
eval mean corners loss: 222.064536
eval box IoU (ground/3D): 0.000000 / 0.000000
eval box estimation accuracy (IoU=0.5): 0.000000
eval box estimation accuracy (IoU=0.7): 0.000000
Model saved in file: /home/FVNet/est-kittinet/outputs/20191219100152/model_1000.ckpt

2. The prediction data is almost not correct, corresponding to label.
   In 007480.txt which in ../prediction/data
   Car 0.00 0.00 0.00 607.93 374.00 612.09 374.00 0.38 0.83 1.29 0.09 189.97 235.72 -3.05 0.63
   Car 0.00 0.00 0.00 687.50 374.00 688.83 374.00 0.19 0.92 0.58 36.43 187.48 334.98 -3.07 0.63
   Car 0.00 0.00 0.00 64.25 293.80 65.31 294.06 0.23 0.71 -1.02 -798.89 177.65 1058.02 -0.17 0.67
   Car 0.00 0.00 0.00 1164.05 360.95 1165.13 361.90 0.67 0.87 -0.27 527.90 179.71 686.34 -0.14 0.66
   Car 0.00 0.00 0.00 757.98 374.00 759.19 374.00 0.38 0.93 -0.22 66.77 189.40 323.57 -0.15 0.67
   Car 0.00 0.00 0.00 1197.39 303.84 1198.13 304.03 0.07 0.96 -0.22 782.06 174.33 959.42 -3.12 0.61
   Car 0.00 0.00 0.00 1183.98 283.32 1184.71 283.51 0.17 0.75 -0.59 901.66 173.54 1131.94 -3.04 0.67
   Car 0.00 0.00 0.00 946.28 374.00 947.86 374.00 0.07 0.90 -0.72 257.56 182.00 550.74 -0.12 0.63
   Car 0.00 0.00 0.00 1228.98 317.58 1230.36 317.76 0.03 0.88 -0.92 749.07 174.96 871.65 -3.11 0.60
   Car 0.00 0.00 0.00 1140.36 374.00 1141.87 374.00 0.03 0.90 -0.40 398.94 179.64 541.61 -0.14 0.60
   Car 0.00 0.00 0.00 63.19 358.00 64.36 358.32 0.01 1.11 -0.39 -518.57 176.04 685.48 -0.13 0.60
   Car 0.00 0.00 0.00 1241.00 364.25 1241.00 364.44 -0.01 0.38 -1.48 601.46 175.87 662.69 -0.16 0.60
   Car 0.00 0.00 0.00 81.58 374.00 83.42 374.00 -0.00 1.16 0.41 -328.90 181.27 450.17 -0.12 0.60

The process is successful and not warning.
Are there some operations I ignore?

Why in kitti_dataset.py " lidar_dir = data_dir + "/cropped/" "

In kitti_dataset.py
Line11: lidar_dir = data_dir + "/cropped/"
Line26: pc_velo_path = self.lidar_dir + img_id + ".npy"
Line27: pc_velo = np.load(pc_velo_path)

Should I use np.fromfile() to load KITTI dataset /training/velodyne data which is ".bin" file?
I don't know what the cropped directory contains. Could you tell me, please? Thanks a lot.

Error in mtrand.RandomState.choice

I have successfully run preprocess_car_person.py, and generated the .npy in cropped directory.
But while I run train.py, it still has a problem like this:

/root/anaconda3/lib/python3.6/site-packages/numpy/core/fromnumeric.py:2957: RuntimeWarning: Mean of empty slice.
out=out, **kwargs)
/root/anaconda3/lib/python3.6/site-packages/numpy/core/_methods.py:73: RuntimeWarning: invalid value encountered in true_divide
ret, rcount, out=ret, casting='unsafe', subok=False)
Traceback (most recent call last):
File "train.py", line 385, in
train()
File "train.py", line 195, in train
train_one_epoch(sess, ops, train_writer)
File "train.py", line 235, in train_one_epoch
get_batch(TRAIN_DATASET, train_idxs, start_idx, end_idx, NUM_CHANNEL)
File "/home/FVNet/est-kittinet/kitti/kitti_dataset.py", line 113, in get_batch
sample = dataset[idxs[(i + start_idx) % num]]
File "/home/FVNet/est-kittinet/kitti/kitti_dataset.py", line 47, in getitem
replace=(object_rect.shape[0] < self.num_point))
File "mtrand.pyx", line 1120, in mtrand.RandomState.choice
ValueError: a must be greater than 0

Do you have met the problem before? Should I normalize some data? Hope for help, thanks.