OA-CNNs training S3DIS and KITTI, as well as nuScenes dataset

[jing-zhao9]

In the OACNNs training code in the readme file, only the training method for training Scannet is seen, and there is no training method for S3DIS, Kitti, or nuScenes. Can you provide a method for training these datasets?

[Xiaoyang Wu]

Hi, getting configs for each other dataset is pretty easy. Just copy our config for SpUNet for each dataset and them replace the model parameters with our provided.

[dzp]

Hi, getting configs for each other dataset is pretty easy. Just copy our config for SpUNet for each dataset and them replace the model parameters with our provided.

Hello, may I ask how to choose the point_grid_size in OACNN for the nuScenes dataset? What parameter did you use to achieve your results?

[Xiaoyang Wu]

Hi, 0.05 for outdoor. Use config for SpUNet as a base, I remember I also add the Lovasz loss.

[dzp]

Hi, 0.05 for outdoor. Use config for SpUNet as a base, I remember I also add the Lovasz loss.

Hi, I am referring to this parameter. point_grid_size=[[8, 12, 16, 16], [6, 9, 12, 12], [4, 6, 8, 8], [3, 4, 6, 6]].

[Xiaoyang Wu]

I remember that it should be the same. (I didn’t have the memory that changed the parameter for the outdoor case)

[jing-zhao9]

@Gofinge

Hi, getting configs for each other dataset is pretty easy. Just copy our config for SpUNet for each dataset and them replace the model parameters with our provided.
This is the config file I set according to what you know above, but it seems that I cannot replicate the miou of S3DIS in the OACNNS paper. I trained OACNNS to replicate the miou in the OACNNS paper based on the config file you provided for scannet. May I ask if my S3DIS configuration file is incorrect. The following is the configuration file for S3DIS

base = ["../base/default_runtime.py"]

misc custom setting

batch_size = 10 # bs: total bs in all gpus
mix_prob = 0.8
empty_cache = False
enable_amp = True

sync_bn = True

sync_bn = False

model settings

model = dict(
type="DefaultSegmentor",
backbone=dict(
type="OACNNs",
in_channels=6,
num_classes=13,
embed_channels=64,
enc_channels=[64, 64, 128, 256],
groups=[4, 4, 8, 16],
enc_depth=[3, 3, 9, 8],
dec_channels=[256, 256, 256, 256],
point_grid_size=[[8, 12, 16, 16], [6, 9, 12, 12], [4, 6, 8, 8], [3, 4, 6, 6]],
dec_depth=[2, 2, 2, 2],
enc_num_ref=[16, 16, 16, 16],
),
criteria=[dict(type="CrossEntropyLoss", loss_weight=1.0, ignore_index=-1)],
)

scheduler settings

epoch = 900
optimizer = dict(type="AdamW", lr=0.001, weight_decay=0.02)
scheduler = dict(
type="OneCycleLR",
max_lr=optimizer["lr"],
pct_start=0.05,
anneal_strategy="cos",
div_factor=10.0,
final_div_factor=1000.0,
)

dataset settings

dataset_type = "S3DISDataset"
data_root = "/home/cvmaster/zj/202404/Pointcept/data/s3dis"

data = dict(
num_classes=13,
ignore_index=-1,
names=[
"ceiling",
"floor",
"wall",
"beam",
"column",
"window",
"door",
"table",
"chair",
"sofa",
"bookcase",
"board",
"clutter",
],
train=dict(
type=dataset_type,
split=("Area_1", "Area_2", "Area_3", "Area_4", "Area_6"),
data_root=data_root,
transform=[
dict(type="CenterShift", apply_z=True),
dict(
type="RandomDropout", dropout_ratio=0.2, dropout_application_ratio=0.2
),
# dict(type="RandomRotateTargetAngle", angle=(1/2, 1, 3/2), center=[0, 0, 0], axis="z", p=0.75),
dict(type="RandomRotate", angle=[-1, 1], axis="z", center=[0, 0, 0], p=0.5),
dict(type="RandomRotate", angle=[-1 / 64, 1 / 64], axis="x", p=0.5),
dict(type="RandomRotate", angle=[-1 / 64, 1 / 64], axis="y", p=0.5),
dict(type="RandomScale", scale=[0.9, 1.1]),
# dict(type="RandomShift", shift=[0.2, 0.2, 0.2]),
dict(type="RandomFlip", p=0.5),
dict(type="RandomJitter", sigma=0.005, clip=0.02),
dict(type="ElasticDistortion", distortion_params=[[0.2, 0.4], [0.8, 1.6]]),
dict(type="ChromaticAutoContrast", p=0.2, blend_factor=None),
dict(type="ChromaticTranslation", p=0.95, ratio=0.05),
dict(type="ChromaticJitter", p=0.95, std=0.05),
# dict(type="HueSaturationTranslation", hue_max=0.2, saturation_max=0.2),
# dict(type="RandomColorDrop", p=0.2, color_augment=0.0),
dict(
type="GridSample",
grid_size=0.05,
hash_type="fnv",
mode="train",
keys=("coord", "color", "segment"),
return_grid_coord=True,
),
dict(type="SphereCrop", point_max=100000, mode="random"),
dict(type="CenterShift", apply_z=False),
dict(type="NormalizeColor"),
dict(type="ShufflePoint"),
dict(type="ToTensor"),
dict(
type="Collect",
keys=("coord", "grid_coord", "segment"),
feat_keys=["coord", "color"],
),
],
test_mode=False,
),
val=dict(
type=dataset_type,
split="Area_5",
data_root=data_root,
transform=[
dict(type="CenterShift", apply_z=True),
dict(
type="GridSample",
grid_size=0.05,
hash_type="fnv",
mode="train",
keys=("coord", "color", "segment"),
return_grid_coord=True,
),
dict(type="CenterShift", apply_z=False),
dict(type="NormalizeColor"),
dict(type="ToTensor"),
dict(
type="Collect",
keys=("coord", "grid_coord", "segment"),
feat_keys=["coord", "color"],
),
],
test_mode=False,
),
test=dict(
type=dataset_type,
split="Area_5",
data_root=data_root,
transform=[dict(type="CenterShift", apply_z=True), dict(type="NormalizeColor")],
test_mode=True,
test_cfg=dict(
voxelize=dict(
type="GridSample",
grid_size=0.05,
hash_type="fnv",
mode="test",
keys=("coord", "color"),
return_grid_coord=True,
),
crop=None,
post_transform=[
dict(type="CenterShift", apply_z=False),
dict(type="ToTensor"),
dict(
type="Collect",
keys=("coord", "grid_coord", "index"),
feat_keys=("coord", "color"),
),
],
aug_transform=[
[dict(type="RandomScale", scale=[0.9, 0.9])],
[dict(type="RandomScale", scale=[0.95, 0.95])],
[dict(type="RandomScale", scale=[1, 1])],
[dict(type="RandomScale", scale=[1.05, 1.05])],
[dict(type="RandomScale", scale=[1.1, 1.1])],
[
dict(type="RandomScale", scale=[0.9, 0.9]),
dict(type="RandomFlip", p=1),
],
[
dict(type="RandomScale", scale=[0.95, 0.95]),
dict(type="RandomFlip", p=1),
],
[dict(type="RandomScale", scale=[1, 1]), dict(type="RandomFlip", p=1)],
[
dict(type="RandomScale", scale=[1.05, 1.05]),
dict(type="RandomFlip", p=1),
],
[
dict(type="RandomScale", scale=[1.1, 1.1]),
dict(type="RandomFlip", p=1),
],
],
),
),
)