Hi, Dear Ph.D Zhao. I have some doubts about details and look forward to your answers

1. I tried generating and submitting test sequences from your pre-trained model to the official codalab. The end result is as follows, which is similar to the result I expected. In fact, in my similar submissions in the past, the test sequence usually has a slightly lower miou than the validation set sequence. So I would like to know if the 60miou you got in your paper was obtained by TTA (test augmentation or fine tuning with validation set, etc.)? In addition, can you provide detailed data（iou class by class） for 60miou？

    ********************************************************************************
    INTERFACE:
    Data:  /tmp/codalab/tmpjFJIFr/run/input/ref
    Predictions:  /tmp/codalab/tmpjFJIFr/run/input/res
    Backend:  numpy
    Split:  test
    Config:  /tmp/codalab/tmpjFJIFr/run/program/semantic-kitti.yaml
    Limit:  None
    Codalab:  /tmp/codalab/tmpjFJIFr/run/output
    ********************************************************************************
    Opening data config file /tmp/codalab/tmpjFJIFr/run/program/semantic-kitti.yaml
    Ignoring xentropy class  0  in IoU evaluation
    [IOU EVAL] IGNORE:  [0]
    [IOU EVAL] INCLUDE:  [ 1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19]
    Evaluating sequences: 10% 20% 30% 40% 50% 60% 70% 80% 90% Validation set:
    Acc avg 0.896
    IoU avg 0.586
    IoU class 1 [car] = 0.930
    IoU class 2 [bicycle] = 0.457
    IoU class 3 [motorcycle] = 0.420
    IoU class 4 [truck] = 0.279
    IoU class 5 [other-vehicle] = 0.326
    IoU class 6 [person] = 0.626
    IoU class 7 [bicyclist] = 0.581
    IoU class 8 [motorcyclist] = 0.305
    IoU class 9 [road] = 0.908
    IoU class 10 [parking] = 0.583
    IoU class 11 [sidewalk] = 0.749
    IoU class 12 [other-ground] = 0.201
    IoU class 13 [building] = 0.885
    IoU class 14 [fence] = 0.595
    IoU class 15 [vegetation] = 0.831
    IoU class 16 [trunk] = 0.643
    IoU class 17 [terrain] = 0.678
    IoU class 18 [pole] = 0.526
    IoU class 19 [traffic-sign] = 0.600
    ********************************************************************************
    below can be copied straight for paper table
    0.930,0.457,0.420,0.279,0.326,0.626,0.581,0.305,0.908,0.583,0.749,0.201,0.885,0.595,0.831,0.643,0.678,0.526,0.600

2. In addition, I still have doubts about the FPS test on SemanticKITTI on all the other works, as there is a lack of uniform and fair comparison(same set of hardware). So I would like to make sure that in the code here, it is necessary to set the sync：torch.cuda.synchronize() https://github.com/placeforyiming/IROS21-FIDNet-SemanticKITTI/blob/7f90b45a765b8bba042b25f642cf12d8fccb5bc2/semantic_inference.py#L170-L173 Just like in RangeNet++. https://github.com/PRBonn/lidar-bonnetal/blob/5a5f4b180117b08879ec97a3a05a3838bce6bb0f/train/tasks/semantic/modules/user.py#L136-L141

Looking forward to your reply again.

Compare with lr_policy 2(Top), lr_policy(Bottom) 1 look like very strange. Kind of like multi warm_up. With lr_policy 1, the loss will change jumping up and down between epochs.

• lr_policy 2:[0.0004, 0.0004, 0.0008, 0.0008, 0.0012, 0.0012, 0.0016, 0.0016, 0.002, 0.002, 0.0018, 0.0018, 0.0016, 0.0016, 0.0014, 0.0014, 0.0012, 0.0012, 0.001, 0.001, 0.0008, 0.0008, 0.0006, 0.0006, 0.0004, 0.0004]

• lr_policy 1:[0.0004, 0.0004, 0.0008, 0.0008, 0.0012, 0.0012, 0.0016, 0.0016, 0.002, 0.002, 0.0016, 0.0016, 0.0012, 0.0012, 0.0008, 0.0008, 0.0012, 0.0012, 0.0006, 0.0006, 0.0008, 0.0008, 0.0004, 0.0004, 0.0002, 0.0002]

hey! I have just discovered your work and I am very interested, I would like to inference with your weights on some of my pointckouds. Given that I have never used a docker I would like to ask you what procedure to follow and what to install to prepare the development environment (or i can directly use the python script without following the previous steps and only installing the requirements library?)(Windows 10 or google collab), thanks

EDIT:I managed to get the python scripts to work, the inference works and the gpu is used correctly. Unfortunately this network has very low performance on my pointclouds, in practice the network assigns label 81 (road sign) to almost all points (obtained with Carla Simulator), I had the same problem also using SalsaNext, SqueezeSegV3 and Darknet53 (this' last is the one that works best as it doesn't put the same id at all points, even if it still has poor performance). The problem from what I understand lies in the geomteria of the pointclouds, I set the simulator parameters to simulate a velodyne hdl64e (the same used to create semanticKITTI) and in fact the point clouds are similar, but despite this the networks are wrong sensationally to make inferences on these points

This code is about ohem? Is there some ablation experiment showing result?https://github.com/placeforyiming/IROS21-FIDNet-SemanticKITTI/blob/7f90b45a765b8bba042b25f642cf12d8fccb5bc2/semantic_main.py#L180-L185