Some questions after running val.py about fvnet

[LordLiang]

maybe the code is too old and not compatible with new preprocessing code. I will check it in time.------------------ 原始邮件 ------------------ 发件人: "hanshong"<[email protected]> 发送时间: 2019年12月20日(星期五) 下午3:21 收件人: "LordLiang/FVNet"<[email protected]>; 抄送: "Subscribed"<[email protected]>; 主题: [LordLiang/FVNet] Some questions after running val.py (#5) 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 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? &mdash; You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[hanshong]

Thanks for your help!
After updating kitti_dataset.py，the results have gained as follow:

eval mean total loss: 8.285833
eval mean center loss: 1.080701
eval mean stage1 center loss: 1.231584
eval mean angle class loss: 0.575514
eval mean angle res loss: 0.008822
eval mean size res loss: 0.245833
eval mean corners loss: 0.060988
eval box IoU (ground/3D): 0.819250 / 0.764644
eval box estimation accuracy (IoU=0.5): 0.962571
eval box estimation accuracy (IoU=0.7): 0.772230

Why AP=77.22 which is obviously higher than your paper's result AP=65.43