Merge pull request #54 from thangvubk/stpls3d

support stpls3d dataset
2025-10-16 11:45:42 +00:00 · 2022-06-28 00:43:20 -05:00 · 2022-06-28 00:43:20 -05:00 · fc1dcbca82
commit fc1dcbca82
parent 737123a7ee 9337a4a69d
13 changed files with 472 additions and 15 deletions
--- a/.gitignore
+++ b/.gitignore
@ -76,3 +76,7 @@ dataset/s3dis/preprocess
 dataset/s3dis/val_gt
 dataset/s3dis/preprocess_sample
 dataset/s3dis/Stanford3dDataset_v1.2
+
+dataset/stpls3d/train
+dataset/stpls3d/val
+dataset/stpls3d/Synthetic_v3_InstanceSegmentation
--- a/configs/softgroup_stpls3d.yaml
+++ b/configs/softgroup_stpls3d.yaml
@ -0,0 +1,84 @@
+model:
+  channels: 16
+  num_blocks: 7
+  semantic_classes: 15
+  instance_classes: 14
+  sem2ins_classes: []
+  semantic_only: False
+  semantic_weight: [1.0, 1.0, 44.0, 21.9, 1.8, 25.1, 31.5, 21.8, 24.0, 54.4, 114.4,
+                    81.2, 43.6, 9.7, 22.4]
+  ignore_label: -100
+  with_coords: False
+  grouping_cfg:
+    score_thr: 0.2
+    radius: 0.9
+    mean_active: 3
+    class_numpoint_mean: [-1., 10408., 58., 124., 1351., 162., 430., 1090., 451., 26., 43.,
+                          61., 39., 109., 1239]
+    npoint_thr: 0.05  # absolute if class_numpoint == -1, relative if class_numpoint != -1
+    ignore_classes: [0]
+  instance_voxel_cfg:
+    scale: 3
+    spatial_shape: 20
+  train_cfg:
+    max_proposal_num: 300
+    pos_iou_thr: 0.5
+    match_low_quality: True
+    min_pos_thr: 0.1
+  test_cfg:
+    x4_split: False
+    cls_score_thr: 0.001
+    mask_score_thr: -0.5
+    min_npoint: 15
+  fixed_modules: []
+
+data:
+  train:
+    type: 'stpls3d'
+    data_root: 'dataset/stpls3d'
+    prefix: 'train'
+    suffix: '_inst_nostuff.pth'
+    training: True
+    repeat: 4
+    voxel_cfg:
+      scale: 3
+      spatial_shape: [128, 512]
+      max_npoint: 250000
+      min_npoint: 5000
+  test:
+    type: 'stpls3d'
+    data_root: 'dataset/stpls3d'
+    prefix: 'val'
+    suffix: '_inst_nostuff.pth'
+    training: False
+    voxel_cfg:
+      scale: 3
+      spatial_shape: [128, 512]
+      max_npoint: 250000
+      min_npoint: 5000
+
+dataloader:
+  train:
+    batch_size: 4
+    num_workers: 4
+  test:
+    batch_size: 1
+    num_workers: 1
+
+optimizer:
+  type: 'Adam'
+  lr: 0.004
+
+save_cfg:
+  semantic: True
+  offset: True
+  instance: True
+
+eval_min_npoint: 10
+
+fp16: False
+epochs: 108
+step_epoch: 20
+save_freq: 4
+pretrain: './work_dirs/softgroup_stpls3d_backbone/latest.pth'
+work_dir: ''
--- a/configs/softgroup_stpls3d_backbone.yaml
+++ b/configs/softgroup_stpls3d_backbone.yaml
@ -0,0 +1,80 @@
+model:
+  channels: 16
+  num_blocks: 7
+  semantic_classes: 15
+  instance_classes: 14
+  sem2ins_classes: []
+  semantic_only: True
+  semantic_weight: [1.0, 1.0, 44.0, 21.9, 1.8, 25.1, 31.5, 21.8, 24.0, 54.4, 114.4,
+                    81.2, 43.6, 9.7, 22.4]
+  with_coords: False
+  ignore_label: -100
+  grouping_cfg:
+    score_thr: 0.2
+    radius: 0.9
+    mean_active: 3
+    class_numpoint_mean: [-1., 10408., 58., 124., 1351., 162., 430., 1090., 451., 26., 43.,
+                          61., 39., 109., 1239]
+    npoint_thr: 0.05  # absolute if class_numpoint == -1, relative if class_numpoint != -1
+    ignore_classes: [0]
+  instance_voxel_cfg:
+    scale: 3
+    spatial_shape: 20
+  train_cfg:
+    max_proposal_num: 200
+    pos_iou_thr: 0.5
+  test_cfg:
+    x4_split: False
+    cls_score_thr: 0.001
+    mask_score_thr: -0.5
+    min_npoint: 100
+  fixed_modules: []
+
+data:
+  train:
+    type: 'stpls3d'
+    data_root: 'dataset/stpls3d'
+    prefix: 'train'
+    suffix: '_inst_nostuff.pth'
+    training: True
+    repeat: 4
+    voxel_cfg:
+      scale: 3
+      spatial_shape: [128, 512]
+      max_npoint: 250000
+      min_npoint: 5000
+  test:
+    type: 'stpls3d'
+    data_root: 'dataset/stpls3d'
+    prefix: 'val'
+    suffix: '_inst_nostuff.pth'
+    training: False
+    voxel_cfg:
+      scale: 3
+      spatial_shape: [128, 512]
+      max_npoint: 250000
+      min_npoint: 5000
+
+dataloader:
+  train:
+    batch_size: 4
+    num_workers: 4
+  test:
+    batch_size: 1
+    num_workers: 1
+
+optimizer:
+  type: 'Adam'
+  lr: 0.004
+
+save_cfg:
+  semantic: True
+  offset: True
+  instance: True
+
+fp16: False
+epochs: 20
+step_epoch: 20
+save_freq: 4
+pretrain: ''
+work_dir: ''
--- a/dataset/stpls3d/prepare_data.sh
+++ b/dataset/stpls3d/prepare_data.sh
@ -0,0 +1,3 @@
+#!/bin/bash
+echo Preprocess data
+python prepare_data_inst_instance_stpls3d.py
--- a/dataset/stpls3d/prepare_data_inst_instance_stpls3d.py
+++ b/dataset/stpls3d/prepare_data_inst_instance_stpls3d.py
@ -0,0 +1,170 @@
+# https://github.com/meidachen/STPLS3D/blob/main/HAIS/data/prepare_data_inst_instance_stpls3d.py
+import glob
+import json
+import math
+import os
+import random
+
+import numpy as np
+import pandas as pd
+import torch
+
+
+def splitPointCloud(cloud, size=50.0, stride=50):
+    limitMax = np.amax(cloud[:, 0:3], axis=0)
+    width = int(np.ceil((limitMax[0] - size) / stride)) + 1
+    depth = int(np.ceil((limitMax[1] - size) / stride)) + 1
+    cells = [(x * stride, y * stride) for x in range(width) for y in range(depth)]
+    blocks = []
+    for (x, y) in cells:
+        xcond = (cloud[:, 0] <= x + size) & (cloud[:, 0] >= x)
+        ycond = (cloud[:, 1] <= y + size) & (cloud[:, 1] >= y)
+        cond = xcond & ycond
+        block = cloud[cond, :]
+        blocks.append(block)
+    return blocks
+
+
+def getFiles(files, fileSplit):
+    res = []
+    for filePath in files:
+        name = os.path.basename(filePath)
+        num = name[:2] if name[:2].isdigit() else name[:1]
+        if int(num) in fileSplit:
+            res.append(filePath)
+    return res
+
+
+def dataAug(file, semanticKeep):
+    points = pd.read_csv(file, header=None).values
+    angle = random.randint(1, 359)
+    angleRadians = math.radians(angle)
+    rotationMatrix = np.array([[math.cos(angleRadians), -math.sin(angleRadians), 0],
+                               [math.sin(angleRadians),
+                                math.cos(angleRadians), 0], [0, 0, 1]])
+    points[:, :3] = points[:, :3].dot(rotationMatrix)
+    pointsKept = points[np.in1d(points[:, 6], semanticKeep)]
+    return pointsKept
+
+
+def preparePthFiles(files, split, outPutFolder, AugTimes=0):
+    # save the coordinates so that we can merge the data to a single scene
+    # after segmentation for visualization
+    outJsonPath = os.path.join(outPutFolder, 'coordShift.json')
+    coordShift = {}
+    # used to increase z range if it is smaller than this,
+    # over come the issue where spconv may crash for voxlization.
+    zThreshold = 6
+
+    # Map relevant classes to {1,...,14}, and ignored classes to -100
+    remapper = np.ones(150) * (-100)
+    for i, x in enumerate([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]):
+        remapper[x] = i
+    # Map instance to -100 based on selected semantic
+    # (change a semantic to -100 if you want to ignore it for instance)
+    remapper_disableInstanceBySemantic = np.ones(150) * (-100)
+    for i, x in enumerate([-100, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]):
+        remapper_disableInstanceBySemantic[x] = i
+
+    # only augment data for these classes
+    semanticKeep = [0, 2, 3, 7, 8, 9, 12, 13]
+
+    counter = 0
+    for file in files:
+
+        for AugTime in range(AugTimes + 1):
+            if AugTime == 0:
+                points = pd.read_csv(file, header=None).values
+            else:
+                points = dataAug(file, semanticKeep)
+            name = os.path.basename(file).strip('.txt') + '_%d' % AugTime
+
+            if split != 'test':
+                coordShift['globalShift'] = list(points[:, :3].min(0))
+            points[:, :3] = points[:, :3] - points[:, :3].min(0)
+
+            blocks = splitPointCloud(points, size=50, stride=50)
+            for blockNum, block in enumerate(blocks):
+                if (len(block) > 10000):
+                    outFilePath = os.path.join(outPutFolder,
+                                               name + str(blockNum) + '_inst_nostuff.pth')
+                    if (block[:, 2].max(0) - block[:, 2].min(0) < zThreshold):
+                        block = np.append(
+                            block, [[
+                                block[:, 0].mean(0), block[:, 1].mean(0), block[:, 2].max(0) +
+                                (zThreshold -
+                                 (block[:, 2].max(0) - block[:, 2].min(0))), block[:, 3].mean(0),
+                                block[:, 4].mean(0), block[:, 5].mean(0), -100, -100
+                            ]],
+                            axis=0)
+                        print('range z is smaller than threshold ')
+                        print(name + str(blockNum) + '_inst_nostuff')
+                    if split != 'test':
+                        outFileName = name + str(blockNum) + '_inst_nostuff'
+                        coordShift[outFileName] = list(block[:, :3].mean(0))
+                    coords = np.ascontiguousarray(block[:, :3] - block[:, :3].mean(0))
+
+                    # coords = block[:, :3]
+                    colors = np.ascontiguousarray(block[:, 3:6]) / 127.5 - 1
+
+                    coords = np.float32(coords)
+                    colors = np.float32(colors)
+                    if split != 'test':
+                        sem_labels = np.ascontiguousarray(block[:, -2])
+                        sem_labels = sem_labels.astype(np.int32)
+                        sem_labels = remapper[np.array(sem_labels)]
+
+                        instance_labels = np.ascontiguousarray(block[:, -1])
+                        instance_labels = instance_labels.astype(np.float32)
+
+                        disableInstanceBySemantic_labels = np.ascontiguousarray(block[:, -2])
+                        disableInstanceBySemantic_labels = disableInstanceBySemantic_labels.astype(
+                            np.int32)
+                        disableInstanceBySemantic_labels = remapper_disableInstanceBySemantic[
+                            np.array(disableInstanceBySemantic_labels)]
+                        instance_labels = np.where(disableInstanceBySemantic_labels == -100, -100,
+                                                   instance_labels)
+
+                        # map instance from 0.
+                        # [1:] because there are -100
+                        uniqueInstances = (np.unique(instance_labels))[1:].astype(np.int32)
+                        remapper_instance = np.ones(50000) * (-100)
+                        for i, j in enumerate(uniqueInstances):
+                            remapper_instance[j] = i
+
+                        instance_labels = remapper_instance[instance_labels.astype(np.int32)]
+
+                        uniqueSemantics = (np.unique(sem_labels))[1:].astype(np.int32)
+
+                        if split == 'train' and (len(uniqueInstances) < 10 or
+                                                 (len(uniqueSemantics) >=
+                                                  (len(uniqueInstances) - 2))):
+                            print('unique insance: %d' % len(uniqueInstances))
+                            print('unique semantic: %d' % len(uniqueSemantics))
+                            print()
+                            counter += 1
+                        else:
+                            torch.save((coords, colors, sem_labels, instance_labels), outFilePath)
+                    else:
+                        torch.save((coords, colors), outFilePath)
+    print('Total skipped file :%d' % counter)
+    json.dump(coordShift, open(outJsonPath, 'w'))
+
+
+if __name__ == '__main__':
+    data_folder = 'Synthetic_v3_InstanceSegmentation'
+    filesOri = sorted(glob.glob(data_folder + '/*.txt'))
+
+    trainSplit = [1, 2, 3, 4, 6, 7, 8, 9, 11, 12, 13, 14, 16, 17, 18, 19, 21, 22, 23, 24]
+    trainFiles = getFiles(filesOri, trainSplit)
+    split = 'train'
+    trainOutDir = split
+    os.makedirs(trainOutDir, exist_ok=True)
+    preparePthFiles(trainFiles, split, trainOutDir, AugTimes=6)
+
+    valSplit = [5, 10, 15, 20, 25]
+    split = 'val'
+    valFiles = getFiles(filesOri, valSplit)
+    valOutDir = split
+    os.makedirs(valOutDir, exist_ok=True)
+    preparePthFiles(valFiles, split, valOutDir)
--- a/dataset/stpls3d/prepare_data_statistic_stpls3d.py
+++ b/dataset/stpls3d/prepare_data_statistic_stpls3d.py
@ -0,0 +1,67 @@
+import glob
+import math
+import os
+
+import numpy as np
+import torch
+
+data_folder = os.path.join(
+    os.path.dirname(os.getcwd()), 'dataset', 'Synthetic_v3_InstanceSegmentation', 'train')
+files = sorted(glob.glob(data_folder + '/*.pth'))
+numclass = 15
+semanticIDs = []
+for i in range(numclass):
+    semanticIDs.append(i)
+
+class_numpoint_mean_dict = {}
+class_radius_mean = {}
+for semanticID in semanticIDs:
+    class_numpoint_mean_dict[semanticID] = []
+    class_radius_mean[semanticID] = []
+num_points_semantic = np.array([0 for i in range(numclass)])
+
+for file in files:
+    coords, colors, sem_labels, instance_labels = torch.load(file)
+    points = np.concatenate(
+        [coords, colors, sem_labels[:, None].astype(int), instance_labels[:, None].astype(int)],
+        axis=1)
+    for semanticID in semanticIDs:
+        singleSemantic = points[np.where(points[:, 6] == semanticID)]
+        uniqueInstances, counts = np.unique(singleSemantic[:, 7], return_counts=True)
+        for count in counts:
+            class_numpoint_mean_dict[semanticID].append(count)
+        allRadius = []
+        for uniqueInstance in uniqueInstances:
+            eachInstance = singleSemantic[np.where(singleSemantic[:, 7] == uniqueInstance)]
+            radius = (np.max(eachInstance, axis=0) - np.min(eachInstance, axis=0)) / 2
+            radius = math.sqrt(radius[0]**2 + radius[1]**2 + radius[2]**2)
+            class_radius_mean[semanticID].append(radius)
+
+    uniqueSemantic, semanticCount = np.unique(points[:, 6], return_counts=True)
+    uniqueSemanticCount = np.array([0 for i in range(numclass)])
+    uniqueSemantic = uniqueSemantic.astype(int)
+    indexOf100 = np.where(uniqueSemantic == -100)
+    semanticCount = np.delete(semanticCount, indexOf100)
+    uniqueSemantic = np.delete(uniqueSemantic, indexOf100)
+    uniqueSemanticCount[uniqueSemantic] = semanticCount
+    num_points_semantic += uniqueSemanticCount
+
+class_numpoint_mean_list = []
+class_radius_mean_list = []
+for semanticID in semanticIDs:
+    class_numpoint_mean_list.append(
+        sum(class_numpoint_mean_dict[semanticID]) * 1.0 / len(class_numpoint_mean_dict[semanticID]))
+    class_radius_mean_list.append(
+        sum(class_radius_mean[semanticID]) / len(class_radius_mean[semanticID]))
+
+print('Using the printed list in hierarchical_aggregation.cpp for class_numpoint_mean_dict: ')
+print([1.0] + [float('{0:0.0f}'.format(i)) for i in class_numpoint_mean_list][1:], sep=',')
+print('Using the printed list in hierarchical_aggregation.cu for class_radius_mean: ')
+print([1.0] + [float('{0:0.2f}'.format(i)) for i in class_radius_mean_list][1:], sep='')
+
+# make ground to 1 the make building to 1
+maxSemantic = np.max(num_points_semantic)
+num_points_semantic = maxSemantic / num_points_semantic
+num_points_semantic = num_points_semantic / num_points_semantic[1]
+print('Using the printed list in hais_run_stpls3d.yaml for class_weight')
+print([1.0, 1.0] + [float('{0:0.2f}'.format(i)) for i in num_points_semantic][2:], sep='')
--- a/softgroup/data/init.py
+++ b/softgroup/data/init.py
@ -3,6 +3,7 @@ from torch.utils.data.distributed import DistributedSampler

 from .s3dis import S3DISDataset
 from .scannetv2 import ScanNetDataset
+from .stpls3d import STPLS3DDataset

 __all__ = ['S3DISDataset', 'ScanNetDataset', 'build_dataset']

@ -16,6 +17,8 @@ def build_dataset(data_cfg, logger):
        return S3DISDataset(**_data_cfg)
    elif data_type == 'scannetv2':
        return ScanNetDataset(**_data_cfg)
+    elif data_type == 'stpls3d':
+        return STPLS3DDataset(**_data_cfg)
    else:
        raise ValueError(f'Unknown {data_type}')

--- a/softgroup/data/custom.py
+++ b/softgroup/data/custom.py
@ -132,9 +132,8 @@ class CustomDataset(Dataset):
        xyz_middle = self.dataAugment(xyz, True, True, True, aug_prob)
        xyz = xyz_middle * self.voxel_cfg.scale
        if np.random.rand() < aug_prob:
-            xyz = self.elastic(xyz, 6 * self.voxel_cfg.scale // 50, 40 * self.voxel_cfg.scale / 50)
-            xyz = self.elastic(xyz, 20 * self.voxel_cfg.scale // 50,
-                               160 * self.voxel_cfg.scale / 50)
+            xyz = self.elastic(xyz, 6, 40.)
+            xyz = self.elastic(xyz, 20, 160.)
        # xyz_middle = xyz / self.voxel_cfg.scale
        xyz = xyz - xyz.min(0)
        max_tries = 5
--- a/softgroup/data/stpls3d.py
+++ b/softgroup/data/stpls3d.py
@ -0,0 +1,15 @@
+from .custom import CustomDataset
+
+
+class STPLS3DDataset(CustomDataset):
+
+    CLASSES = ('building', 'low vegetation', 'med. vegetation', 'high vegetation', 'vehicle',
+               'truck', 'aircraft', 'militaryVehicle', 'bike', 'motorcycle', 'light pole',
+               'street sign', 'clutter', 'fence')
+
+    def getInstanceInfo(self, xyz, instance_label, semantic_label):
+        ret = super().getInstanceInfo(xyz, instance_label, semantic_label)
+        instance_num, instance_pointnum, instance_cls, pt_offset_label = ret
+        # ignore instance of class 0 and reorder class id
+        instance_cls = [x - 1 if x != -100 else x for x in instance_cls]
+        return instance_num, instance_pointnum, instance_cls, pt_offset_label
--- a/softgroup/evaluation/instance_eval.py
+++ b/softgroup/evaluation/instance_eval.py
@ -12,7 +12,7 @@ from .instance_eval_util import get_instances

 class ScanNetEval(object):

-    def __init__(self, class_labels, iou_type=None, use_label=True):
+    def __init__(self, class_labels, min_npoint=None, iou_type=None, use_label=True):
        self.valid_class_labels = class_labels
        self.valid_class_ids = np.arange(len(class_labels)) + 1
        self.id2label = {}
@ -22,7 +22,10 @@ class ScanNetEval(object):
            self.id2label[self.valid_class_ids[i]] = self.valid_class_labels[i]

        self.ious = np.append(np.arange(0.5, 0.95, 0.05), 0.25)
-        self.min_region_sizes = np.array([100])
+        if min_npoint:
+            self.min_region_sizes = np.array([min_npoint])
+        else:
+            self.min_region_sizes = np.array([100])
        self.distance_threshes = np.array([float('inf')])
        self.distance_confs = np.array([-float('inf')])

--- a/softgroup/model/softgroup.py
+++ b/softgroup/model/softgroup.py
@ -21,8 +21,10 @@ class SoftGroup(nn.Module):
                 semantic_only=False,
                 semantic_classes=20,
                 instance_classes=18,
+                 semantic_weight=None,
                 sem2ins_classes=[],
                 ignore_label=-100,
+                 with_coords=True,
                 grouping_cfg=None,
                 instance_voxel_cfg=None,
                 train_cfg=None,
@ -34,8 +36,10 @@ class SoftGroup(nn.Module):
        self.semantic_only = semantic_only
        self.semantic_classes = semantic_classes
        self.instance_classes = instance_classes
+        self.semantic_weight = semantic_weight
        self.sem2ins_classes = sem2ins_classes
        self.ignore_label = ignore_label
+        self.with_coords = with_coords
        self.grouping_cfg = grouping_cfg
        self.instance_voxel_cfg = instance_voxel_cfg
        self.train_cfg = train_cfg
@ -46,9 +50,10 @@ class SoftGroup(nn.Module):
        norm_fn = functools.partial(nn.BatchNorm1d, eps=1e-4, momentum=0.1)

        # backbone
+        in_channels = 6 if with_coords else 3
        self.input_conv = spconv.SparseSequential(
            spconv.SubMConv3d(
-                6, channels, kernel_size=3, padding=1, bias=False, indice_key='subm1'))
+                in_channels, channels, kernel_size=3, padding=1, bias=False, indice_key='subm1'))
        block_channels = [channels * (i + 1) for i in range(num_blocks)]
        self.unet = UBlock(block_channels, norm_fn, 2, block, indice_key_id=1)
        self.output_layer = spconv.SparseSequential(norm_fn(channels), nn.ReLU())
@ -103,7 +108,8 @@ class SoftGroup(nn.Module):
                      semantic_labels, instance_labels, instance_pointnum, instance_cls,
                      pt_offset_labels, spatial_shape, batch_size, **kwargs):
        losses = {}
-        feats = torch.cat((feats, coords_float), 1)
+        if self.with_coords:
+            feats = torch.cat((feats, coords_float), 1)
        voxel_feats = voxelization(feats, p2v_map)
        input = spconv.SparseConvTensor(voxel_feats, voxel_coords.int(), spatial_shape, batch_size)
        semantic_scores, pt_offsets, output_feats = self.forward_backbone(input, v2p_map)
@ -140,8 +146,12 @@ class SoftGroup(nn.Module):
    def point_wise_loss(self, semantic_scores, pt_offsets, semantic_labels, instance_labels,
                        pt_offset_labels):
        losses = {}
+        if self.semantic_weight:
+            weight = torch.tensor(self.semantic_weight, dtype=torch.float, device='cuda')
+        else:
+            weight = None
        semantic_loss = F.cross_entropy(
-            semantic_scores, semantic_labels, ignore_index=self.ignore_label)
+            semantic_scores, semantic_labels, weight=weight, ignore_index=self.ignore_label)
        losses['semantic_loss'] = semantic_loss

        pos_inds = instance_labels != self.ignore_label
@ -169,14 +179,30 @@ class SoftGroup(nn.Module):
        fg_instance_cls = instance_cls[fg_inds]
        fg_ious_on_cluster = ious_on_cluster[:, fg_inds]

+        # assign proposal to gt idx. -1: negative, 0 -> num_gts - 1: positive
+        num_proposals = fg_ious_on_cluster.size(0)
+        num_gts = fg_ious_on_cluster.size(1)
+        assigned_gt_inds = fg_ious_on_cluster.new_full((num_proposals, ), -1, dtype=torch.long)
+
        # overlap > thr on fg instances are positive samples
-        max_iou, gt_inds = fg_ious_on_cluster.max(1)
+        max_iou, argmax_iou = fg_ious_on_cluster.max(1)
        pos_inds = max_iou >= self.train_cfg.pos_iou_thr
-        pos_gt_inds = gt_inds[pos_inds]
+        assigned_gt_inds[pos_inds] = argmax_iou[pos_inds]
+
+        # allow low-quality proposals with best iou to be as positive sample
+        # in case pos_iou_thr is too high to achieve
+        match_low_quality = getattr(self.train_cfg, 'match_low_quality', False)
+        min_pos_thr = getattr(self.train_cfg, 'min_pos_thr', 0)
+        if match_low_quality:
+            gt_max_iou, gt_argmax_iou = fg_ious_on_cluster.max(0)
+            for i in range(num_gts):
+                if gt_max_iou[i] >= min_pos_thr:
+                    assigned_gt_inds[gt_argmax_iou[i]] = i

        # compute cls loss. follow detection convention: 0 -> K - 1 are fg, K is bg
-        labels = fg_instance_cls.new_full((fg_ious_on_cluster.size(0), ), self.instance_classes)
-        labels[pos_inds] = fg_instance_cls[pos_gt_inds]
+        labels = fg_instance_cls.new_full((num_proposals, ), self.instance_classes)
+        pos_inds = assigned_gt_inds >= 0
+        labels[pos_inds] = fg_instance_cls[assigned_gt_inds[pos_inds]]
        cls_loss = F.cross_entropy(cls_scores, labels)
        losses['cls_loss'] = cls_loss

@ -221,7 +247,8 @@ class SoftGroup(nn.Module):
    def forward_test(self, batch_idxs, voxel_coords, p2v_map, v2p_map, coords_float, feats,
                     semantic_labels, instance_labels, pt_offset_labels, spatial_shape, batch_size,
                     scan_ids, **kwargs):
-        feats = torch.cat((feats, coords_float), 1)
+        if self.with_coords:
+            feats = torch.cat((feats, coords_float), 1)
        voxel_feats = voxelization(feats, p2v_map)
        input = spconv.SparseConvTensor(voxel_feats, voxel_coords.int(), spatial_shape, batch_size)
        semantic_scores, pt_offsets, output_feats = self.forward_backbone(
--- a/tools/test.py
+++ b/tools/test.py
@ -116,7 +116,8 @@ def main():
                gt_insts.append(res['gt_instances'])
        if not cfg.model.semantic_only:
            logger.info('Evaluate instance segmentation')
-            scannet_eval = ScanNetEval(dataset.CLASSES)
+            eval_min_npoint = getattr(cfg, 'eval_min_npoint', None)
+            scannet_eval = ScanNetEval(dataset.CLASSES, eval_min_npoint)
            scannet_eval.evaluate(pred_insts, gt_insts)
        logger.info('Evaluate semantic segmentation and offset MAE')
        ignore_label = cfg.model.ignore_label
--- a/tools/train.py
+++ b/tools/train.py
@ -108,7 +108,8 @@ def validate(epoch, model, val_loader, cfg, logger, writer):
                all_gt_insts.append(res['gt_instances'])
        if not cfg.model.semantic_only:
            logger.info('Evaluate instance segmentation')
-            scannet_eval = ScanNetEval(val_set.CLASSES)
+            eval_min_npoint = getattr(cfg, 'eval_min_npoint', None)
+            scannet_eval = ScanNetEval(val_set.CLASSES, eval_min_npoint)
            eval_res = scannet_eval.evaluate(all_pred_insts, all_gt_insts)
            writer.add_scalar('val/AP', eval_res['all_ap'], epoch)
            writer.add_scalar('val/AP_50', eval_res['all_ap_50%'], epoch)