lidargaitv2 open-source

opengait/data/collate_fn.py

@@ -33,6 +33,8 @@ class CollateFn(object):
         if self.sampler == 'all' and 'frames_all_limit' in sample_config:
             self.frames_all_limit = sample_config['frames_all_limit']
 
+        self.points_in_use = sample_config.get('points_in_use')
+
     def __call__(self, batch):
         batch_size = len(batch)
         # currently, the functionality of feature_num is not fully supported yet, it refers to 1 now. We are supposed to make our framework support multiple source of input data, such as silhouette, or skeleton.
@@ -88,7 +90,14 @@ class CollateFn(object):
 
             for i in range(feature_num):
                 for j in indices[:self.frames_all_limit] if self.frames_all_limit > -1 and len(indices) > self.frames_all_limit else indices:
-                    sampled_fras[i].append(seqs[i][j])
+                    point_cloud_index = self.points_in_use.get('pointcloud_index')
+                    if self.points_in_use is not None and point_cloud_index is not None and i == point_cloud_index:
+                        points_num = self.points_in_use.get('points_num')
+                        sample_points = (random.choices(range(len(seqs[i][j])), k=points_num)
+                                if points_num is not None else list(range(len(seqs[i][j]))))
+                        sampled_fras[i].append(np.asarray([seqs[i][j][p] for p in sample_points]))
+                    else:
+                        sampled_fras[i].append(seqs[i][j])
             return sampled_fras
 
         # f: feature_num
@@ -112,4 +121,4 @@ class CollateFn(object):
             batch[-1] = np.asarray(seqL_batch)
 
         batch[0] = fras_batch
-        return batch
+        return batch

opengait/data/transform.py

@@ -228,6 +228,121 @@ def get_transform(trf_cfg=None):
         return transform
     raise "Error type for -Transform-Cfg-"
 
+# **************** For LidarGait++ ****************
+# Shen, et al: LidarGait++: Learning Local Features and Size Awareness from LiDAR Point Clouds for 3D Gait Recognition, CVPR2025
+
+def normalize_point_cloud(batch_data):
+    """Normalize the batch data using coordinates of the block centered at origin.
+    
+    Input:
+        batch_data: BxNxC array
+    Output:
+        BxNxC array
+    """
+    centroids = np.mean(batch_data, axis=1, keepdims=True)  # shape: (B, 1, C)
+    centered = batch_data - centroids
+    scales = np.max(np.linalg.norm(centered, axis=2), axis=1, keepdims=True)  # shape: (B, 1)
+    scales = scales.reshape(batch_data.shape[0], 1, 1)  # (B, 1, 1) for broadcasting
+    return centered / scales
+
+
+def dropout_point_cloud(batch_data, max_dropout_ratio=0.875, prob=0.2):
+    """Randomly drop points in each point cloud.
+    
+    Input:
+        batch_data: BxNx3 array
+    Output:
+        BxNx3 array, with dropped points replaced by the first point in each cloud.
+    """
+    if np.random.rand() >= prob:
+        return batch_data
+    B, N, C = batch_data.shape
+    # 为每个点云生成一个 dropout_ratio（范围 0 ~ max_dropout_ratio）
+    dropout_ratio = np.random.rand(B, 1) * max_dropout_ratio  # shape: (B, 1)
+    random_matrix = np.random.rand(B, N)
+    drop_mask = random_matrix <= dropout_ratio  # shape: (B, N)
+    # 构造每个点云第一个点重复 N 次的数组，用于替换被 dropout 的点
+    first_points = np.repeat(batch_data[:, :1, :], N, axis=1)
+    return np.where(drop_mask[..., None], first_points, batch_data)
+    
+def shift_point_cloud(batch_data, shift_range=0.1, prob=0.2):
+    """ Randomly shift point cloud. Shift is per point cloud.
+        Input:
+          BxNx3 array, original batch of point clouds
+        Return:
+          BxNx3 array, shifted batch of point clouds
+    """
+    if np.random.rand() >= prob:
+        return batch_data
+    B, N, C = batch_data.shape
+    shifts = np.random.uniform(-shift_range, shift_range, (B, N,3))
+    batch_data += shifts
+    return batch_data
+
+
+def scale_point_cloud(batch_data, scale_low=0.8, scale_high=1.25, prob=0.2):
+    """ Randomly scale the point cloud. Scale is per point cloud.
+        Input:
+            BxNx3 array, original batch of point clouds
+        Return:
+            BxNx3 array, scaled batch of point clouds
+    """
+    if np.random.rand() >= prob:
+        return batch_data
+    B, N, C = batch_data.shape
+    scales = np.random.uniform(scale_low, scale_high, B)
+    for batch_index in range(B):
+        batch_data[batch_index,:,:] *= scales[batch_index]
+    return batch_data
+
+def jitter_point_cloud(batch_data, std=0.01, clip=0.05, prob=0.2):
+    if np.random.rand() >= prob:
+        return batch_data
+    B, N, C = batch_data.shape
+    jittered_data = np.random.normal(loc=0.0, scale=std, size=(B, N, C))
+    jittered_data = np.clip(jittered_data, -clip, clip)
+    batch_data += jittered_data
+    return batch_data
+
+def flip_point_cloud_y(batch_data, prob=0.25):
+    if np.random.rand() >= prob:
+        return batch_data
+    batch_data[:, :, 1] = -batch_data[:, :, 1]
+    return batch_data
+
+
+def getxyz(batch_data,col = 2,to_ground=False):
+    B,N,C = batch_data.shape
+    last_col = batch_data[:, :, col]
+    result = last_col.reshape((B, N, 1))
+    if to_ground:
+        result -= result.min(axis=1,keepdims=True)
+    return result
+
+class PointCloudsTransform():
+    def __init__(self, xyz_only=True, scale_aware=False, drop_prob=0, shift_prob=0, jit_prob=0,scale_prob=0, flip_prob=0):
+        self.scale_aware = scale_aware
+        self.xyz_only = xyz_only
+
+        self.flip_prob, self.shift_prob, self.jit_prob, self.scale_prob, self.drop_prob = flip_prob, shift_prob, jit_prob, scale_prob, drop_prob 
+    def __call__(self, points):
+        if self.xyz_only:
+            points = points[:,:,:3]
+        
+        heights = getxyz(points, col = 2, to_ground=True)
+        points = normalize_point_cloud(points)
+
+        points = flip_point_cloud_y(points, prob=self.flip_prob)
+        points = shift_point_cloud(points, prob=self.shift_prob)
+        points = jitter_point_cloud(points, prob=self.jit_prob)
+        points = scale_point_cloud(points, prob=self.scale_prob)
+        points = dropout_point_cloud(points, prob=self.drop_prob)
+
+        if self.scale_aware:
+            points = np.concatenate([points,heights],axis=-1)
+        return points
+
+
 # **************** For GaitSSB ****************
 # Fan, et al: Learning Gait Representation from Massive Unlabelled Walking Videos: A Benchmark, T-PAMI2023
 
@@ -587,4 +702,4 @@ class MSGGTransform():
         
     def __call__(self, x):
         result=x[...,self.mask,:].copy()
-        return result
+        return result

opengait/evaluation/evaluator.py

@@ -456,4 +456,45 @@ def evaluate_scoliosis(data, dataset, metric='euc'):
         print(f"{cls} Specificity: {TNR[i] * 100:.2f}%")
     print(f"Accuracy: {accuracy * 100:.2f}%")
 
-    return result_dict
+    return result_dict
+
+def evaluate_FreeGait(data, dataset, metric='euc'):
+    msg_mgr = get_msg_mgr()
+
+    features, labels, cams, time_seqs = data['embeddings'], data['labels'], data['types'], data['views']
+    import json
+    probe_sets = json.load(
+        open('./datasets/FreeGait/FreeGait.json', 'rb'))['PROBE_SET']
+    
+    probe_mask = []
+    for id, ty, sq in zip(labels, cams, time_seqs):
+        if '-'.join([id, ty, sq]) in probe_sets:
+            probe_mask.append(True)
+        else:
+            probe_mask.append(False)
+    probe_mask = np.array(probe_mask)
+
+    # probe_features = features[:probe_num]
+    probe_features = features[probe_mask]
+    # gallery_features = features[probe_num:]
+    gallery_features = features[~probe_mask]
+    # probe_lbls = np.asarray(labels[:probe_num])
+    # gallery_lbls = np.asarray(labels[probe_num:])
+    probe_lbls = np.asarray(labels)[probe_mask]
+    gallery_lbls = np.asarray(labels)[~probe_mask]
+
+    results = {}
+    msg_mgr.log_info(f"The test metric you choose is {metric}.")
+    dist = cuda_dist(probe_features, gallery_features, metric).cpu().numpy()
+    cmc, all_AP, all_INP = evaluate_rank(dist, probe_lbls, gallery_lbls)
+
+    mAP = np.mean(all_AP)
+    mINP = np.mean(all_INP)
+    for r in [1, 5, 10]:
+        results['scalar/test_accuracy/Rank-{}'.format(r)] = cmc[r - 1] * 100
+    results['scalar/test_accuracy/mAP'] = mAP * 100
+    results['scalar/test_accuracy/mINP'] = mINP * 100
+
+    # print_csv_format(dataset_name, results)
+    msg_mgr.log_info(results)
+    return results

opengait/modeling/models/lidargaitv2.py

@@ -0,0 +1,83 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+from .lidargaitv2_utils import PointNetSetAbstraction, PPPooling, PPPooling_UDP,NetVLAD
+
+from ..base_model import BaseModel
+from ..modules import SeparateFCs, SeparateBNNecks
+
+
+class LidarGaitPlusPlus(BaseModel):
+    def build_network(self, model_cfg):
+        C = model_cfg['channel']
+        C_out = model_cfg['SeparateFCs']['in_channels']
+        scale_aware = model_cfg['scale_aware']
+        normalize_dp = model_cfg['normalize_dp']
+        sampling = model_cfg['sampling']
+
+        npoints = model_cfg.get('npoints', [512, 256, 128])
+        nsample = model_cfg.get('nsample', 32)
+        in_channel = 4 if scale_aware else 3
+
+        self.sa1 = PointNetSetAbstraction(npoint=npoints[0], radius=0.1, nsample=nsample, in_channel=in_channel, mlp=[2*C, 2*C, 4*C], group_all=False, sampling=sampling, scale_aware=scale_aware, normalize_dp=normalize_dp)
+        self.sa2 = PointNetSetAbstraction(npoint=npoints[1], radius=0.2, nsample=nsample, in_channel=4*C + in_channel, mlp=[4*C, 4*C, 8*C], group_all=False, sampling=sampling, scale_aware=scale_aware, normalize_dp=normalize_dp)
+        self.sa3 = PointNetSetAbstraction(npoint=npoints[2], radius=0.4, nsample=nsample, in_channel=8*C + in_channel, mlp=[8*C, 8*C, 16*C], group_all=False, sampling=sampling, scale_aware=scale_aware, normalize_dp=normalize_dp)
+        self.sa4 = PointNetSetAbstraction(npoint=None, radius=None, nsample=None, in_channel=16*C + in_channel, mlp=[16*C, 16*C, C_out], group_all=True, sampling=sampling, scale_aware=scale_aware, normalize_dp=normalize_dp)
+
+        if model_cfg['pool'] == 'VLAD':
+            self.pool = NetVLAD(num_clusters=16, dim=C_out, alpha=1.0)
+        elif model_cfg['pool'] == 'GMaxP':
+            self.pool = PPPooling_UDP([1])
+        elif model_cfg['pool'] == 'PPP_UDP':
+            self.pool = PPPooling_UDP(model_cfg['scale'])
+        elif model_cfg['pool'] == 'PPP_UAP':
+            self.pool = PPPooling(scale_aware=False, bin_num=model_cfg['scale'])
+        elif model_cfg['pool'] == 'PPP_HAP':
+            self.pool = PPPooling(scale_aware=True, bin_num=model_cfg['scale'])
+        
+
+
+        self.BNNecks = SeparateBNNecks(**model_cfg['SeparateBNNecks'])
+        self.FCs = SeparateFCs(**model_cfg['SeparateFCs']) 
+
+
+    def forward(self, inputs):
+        ipts, labs, _, views, seqL = inputs
+
+        xyz = ipts[0]
+        B, T, N, C = xyz.shape
+        xyz = rearrange(xyz, 'B T N C -> (B T) C N')
+
+
+        l1_xyz, l1_points = self.sa1(xyz, None)
+        l1_points = torch.max(l1_points, dim=-2)[0]
+
+        l2_xyz, l2_points = self.sa2(l1_xyz, l1_points)
+        l2_points = torch.max(l2_points, dim=-2)[0]
+
+        l3_xyz, l3_points = self.sa3(l2_xyz, l2_points)
+        l3_points = torch.max(l3_points, dim=-2)[0]
+
+        l4_xyz, l4_points = self.sa4(l3_xyz, l3_points)
+
+        x = self.pool(l4_points, l3_xyz)
+
+
+        x = rearrange(x, '(B T) feat p -> B T feat p', B=B)
+        feat = x.max(1)[0]# x.mean(1) # x.max(1)[0]
+        embed = self.FCs(feat)  # [n, c, p]
+        embed_2, logits = self.BNNecks(embed)  # [n, c, p]
+        retval = {
+            'training_feat': {
+                'triplet': {'embeddings': embed, 'labels': labs},
+                'softmax': {'logits': logits, 'labels': labs}
+            },
+            'visual_summary': {
+            },
+            'inference_feat': {
+                'embeddings': embed,
+            }
+        }
+        return retval

opengait/modeling/models/lidargaitv2_utils.py

@@ -0,0 +1,377 @@
+import torch.nn as nn
+import torch.nn.functional as F
+import torch
+import torch.nn as nn
+import torch.utils.data
+import torch.nn.functional as F
+from einops import rearrange
+from torch.autograd import Variable
+import numpy as np
+
+
+def square_distance(src, dst):
+    """
+    Calculate Euclid distance between each two points.
+    src^T * dst = xn * xm + yn * ym + zn * zm；
+    sum(src^2, dim=-1) = xn*xn + yn*yn + zn*zn;
+    sum(dst^2, dim=-1) = xm*xm + ym*ym + zm*zm;
+    dist = (xn - xm)^2 + (yn - ym)^2 + (zn - zm)^2
+         = sum(src**2,dim=-1)+sum(dst**2,dim=-1)-2*src^T*dst
+    Input:
+        src: source points, [B, N, C]
+        dst: target points, [B, M, C]
+    Output:
+        dist: per-point square distance, [B, N, M]
+    """
+    B, N, _ = src.shape
+    _, M, _ = dst.shape
+    dist = -2 * torch.matmul(src, dst.permute(0, 2, 1))
+    dist += torch.sum(src ** 2, -1).view(B, N, 1)
+    dist += torch.sum(dst ** 2, -1).view(B, 1, M)
+    return dist
+
+
+def index_points(points, idx):
+    """
+    Input:
+        points: input points data, [B, N, C]
+        idx: sample index data, [B, S]
+    Return:
+        new_points:, indexed points data, [B, S, C]
+    """
+    device = points.device
+    B = points.shape[0]
+    view_shape = list(idx.shape)
+    view_shape[1:] = [1] * (len(view_shape) - 1)
+    repeat_shape = list(idx.shape)
+    repeat_shape[0] = 1
+    batch_indices = torch.arange(B, dtype=torch.long).to(device).view(view_shape).repeat(repeat_shape)
+    new_points = points[batch_indices, idx, :]
+    return new_points
+
+
+def farthest_point_sample(xyz, npoint):
+    """
+    Input:
+        xyz: pointcloud data, [B, N, 3]
+        npoint: number of samples
+    Return:
+        centroids: sampled pointcloud index, [B, npoint]
+    """
+    device = xyz.device
+    B, N, C = xyz.shape
+    centroids = torch.zeros(B, npoint, dtype=torch.long).to(device)
+    distance = torch.ones(B, N).to(device) * 1e10
+    farthest = torch.randint(0, N, (B,), dtype=torch.long).to(device)
+    batch_indices = torch.arange(B, dtype=torch.long).to(device)
+    for i in range(npoint):
+        centroids[:, i] = farthest
+        centroid = xyz[batch_indices, farthest, :].view(B, 1, C)
+        dist = torch.sum((xyz - centroid) ** 2, -1)
+        mask = dist < distance
+        distance[mask] = dist[mask]
+        farthest = torch.max(distance, -1)[1]
+    return centroids
+
+
+def ball_query(radius, nsample, xyz, new_xyz):
+    """
+    Input:
+        radius: local region radius
+        nsample: max sample number in local region
+        xyz: all points, [B, N, 3]
+        new_xyz: query points, [B, S, 3]
+    Return:
+        group_idx: grouped points index, [B, S, nsample]
+    """
+    device = xyz.device
+    B, N, C = xyz.shape
+    _, S, _ = new_xyz.shape
+    group_idx = torch.arange(N, dtype=torch.long).to(device).view(1, 1, N).repeat([B, S, 1])
+    xyz = xyz[:,:,:3]
+    new_xyz = new_xyz[:,:,:3]
+    sqrdists = square_distance(new_xyz, xyz)
+    group_idx[sqrdists > radius ** 2] = N
+    group_idx = group_idx.sort(dim=-1)[0][:, :, :nsample]
+    group_first = group_idx[:, :, 0].view(B, S, 1).repeat([1, 1, nsample])
+    mask = group_idx == N
+    group_idx[mask] = group_first[mask]
+    return group_idx
+
+
+def knn_query(k, xyz, new_xyz):
+    """
+    Input:
+        k: number of nearest neighbors to query
+        xyz: all points, [B, N, 3]
+        new_xyz: query points, [B, S, 3]
+    Return:
+        group_idx: indices of k-nearest neighbors, [B, S, k]
+    """
+    B, N, C = xyz.shape
+    _, S, _ = new_xyz.shape
+    xyz = xyz[:,:,:3]
+    new_xyz = new_xyz[:,:,:3]
+    dists = square_distance(new_xyz, xyz)
+    #scaling_factor = torch.Tensor([1, 1, 0.6]).to(new_xyz.device)
+    #dists = torch.sum(torch.square(new_xyz.unsqueeze(2) - xyz.unsqueeze(1)) / scaling_factor, dim=-1)
+    group_idx = dists.sort(dim=-1)[1][:, :, :k]
+    return group_idx
+
+
+def sample_and_group(npoint, radius, nsample, xyz, points, returnfps=False, sampling='ball',scale_aware=False, normalize_dp=False):
+    """
+    Input:
+        npoint:
+        radius:
+        nsample:
+        xyz: input points position data, [B, N, 3]
+        points: input points data, [B, N, D]
+    Return:
+        new_xyz: sampled points position data, [B, npoint, nsample, 3]
+        new_points: sampled points data, [B, npoint, nsample, 3+D]
+    """
+    B, N, C = xyz.shape
+    S = npoint
+    fps_idx = farthest_point_sample(xyz[:,:,:3], npoint)  # [B, npoint, C]
+    new_xyz = index_points(xyz, fps_idx)
+
+    if sampling == 'ball':
+        idx =  ball_query(radius, nsample, xyz, new_xyz)
+    elif sampling == 'knn':
+        idx =  knn_query(nsample, xyz, new_xyz)
+    else:
+        raise ValueError("Unsupported sampling type. Use 'ball' or 'knn'.")
+
+    grouped_xyz = index_points(xyz, idx) # [B, npoint, nsample, C]
+    grouped_xyz_norm = grouped_xyz - new_xyz.view(B, S, 1, C)
+    if normalize_dp:  # and sampling!='knn':
+        grouped_xyz_norm /= radius
+    grouped_xyz_norm = grouped_xyz_norm if scale_aware else grouped_xyz_norm[:,:,:,:3]
+
+    
+    if points is not None:
+        grouped_points = index_points(points, idx)
+        new_points = torch.cat([grouped_xyz_norm, grouped_points], dim=-1) # [B, npoint, nsample, C+D]
+    else:
+        new_points = grouped_xyz_norm
+    if returnfps:
+        return new_xyz, new_points, grouped_xyz, fps_idx
+    else:
+        return new_xyz, new_points
+
+
+def sample_and_group_all(xyz, points, scale_aware=False):
+    """
+    Input:
+        xyz: input points position data, [B, N, 3]
+        points: input points data, [B, N, D]
+    Return:
+        new_xyz: sampled points position data, [B, 1, 3]
+        new_points: sampled points data, [B, 1, N, 3+D]
+    """
+    device = xyz.device
+    B, N, C = xyz.shape
+    new_xyz = torch.zeros(B, 1, C).to(device)
+    grouped_xyz = xyz.view(B, 1, N, C)
+    grouped_xyz = grouped_xyz if scale_aware else grouped_xyz[:,:,:,:3]
+    if points is not None:
+        new_points = torch.cat([grouped_xyz, points.view(B, 1, N, -1)], dim=-1)
+    else:
+        new_points = grouped_xyz
+    return new_xyz, new_points
+
+
+class PointNetSetAbstraction(nn.Module):
+    def __init__(self, npoint, radius, nsample, in_channel, mlp, group_all, sampling='ball', scale_aware=False,normalize_dp=False):
+        super(PointNetSetAbstraction, self).__init__()
+        self.npoint = npoint
+        self.radius = radius
+        self.nsample = nsample
+        self.mlp_convs = nn.ModuleList()
+        self.mlp_bns = nn.ModuleList()
+        last_channel = in_channel
+        for out_channel in mlp:
+            self.mlp_convs.append(nn.Conv2d(last_channel, out_channel, 1))
+            self.mlp_bns.append(nn.BatchNorm2d(out_channel))
+            last_channel = out_channel
+        self.group_all = group_all
+        self.scale_aware = scale_aware
+        self.normalize_dp = normalize_dp
+        self.sampling = sampling
+    def forward(self, xyz, points):
+        """
+        Input:
+            xyz: input points position data, [B, C, N]
+            points: input points data, [B, D, N]
+        Return:
+            new_xyz: sampled points position data, [B, C, S]
+            new_points_concat: sample points feature data, [B, D', S]
+        """
+        xyz = xyz.permute(0, 2, 1)
+        if points is not None:
+            points = points.permute(0, 2, 1)
+
+        if self.group_all:
+            new_xyz, new_points = sample_and_group_all(xyz, points, scale_aware=self.scale_aware)
+        else:
+            new_xyz, new_points = sample_and_group(self.npoint, self.radius, self.nsample, xyz, points, sampling=self.sampling, scale_aware=self.scale_aware,normalize_dp=self.normalize_dp)
+        # new_xyz: sampled points position data, [B, ], C]
+        # new_points: sampled points data, [B, npoint, nsample, C+D]
+        new_points = new_points.permute(0, 3, 2, 1) # [B, C+D, nsample,npoint]
+        for i, conv in enumerate(self.mlp_convs):
+            bn = self.mlp_bns[i]
+            new_points =  F.relu(bn(conv(new_points)))
+
+        new_points = new_points
+        new_xyz = new_xyz.permute(0, 2, 1)
+        return new_xyz, new_points
+
+class PPPooling_UDP():
+    """
+        Hierarchically Clustered Point Pooling
+    """
+
+    def __init__(self, bin_num=None):
+        if bin_num is None:
+            bin_num = [16, 8, 4, 2, 1]
+        self.bin_num = bin_num
+
+    def __call__(self, x, xyz):
+        """
+            x  : [n, c, h, w]
+            xyz: [n, 3, p]
+            ret: [n, c, p] 
+        """
+        #print(xyz.shape)
+        #x = rearrange(x, 'b n c -> b c n 1')
+        n, c = x.size()[:2]
+        _, idx = xyz[:, 2, :].sort()
+        x = x.gather(2, idx.unsqueeze(1).unsqueeze(-1).expand_as(x))
+        features = []
+        for b in self.bin_num:
+            z = x.view(n, c, b, -1)
+            z = z.mean(-1) + z.max(-1)[0]
+            features.append(z)
+        return torch.cat(features, -1)
+
+
+class PPPooling():
+    def __init__(self, scale_aware=False, bin_num=None):
+        # 默认设置多个分辨率的分bin数量
+        self.bin_num = bin_num if bin_num is not None else [16, 8, 4, 2, 1]
+        self.scale_aware = scale_aware
+
+    def __call__(self, point_clouds, points):
+        # 调整维度：输入 point_clouds: B x C x N x 1 转换为 B x N x C，
+        # points: B x C x N 转换为 B x N x C
+        point_clouds = rearrange(point_clouds, 'B C N 1 -> B N C')
+        points = rearrange(points, 'B C N -> B N C')
+        B, N, C = point_clouds.shape
+
+        if self.scale_aware: # PPPooling_HAP
+            z = points[:, :, 3]  # shape: (B, N)
+            # 固定的 z 范围（例如：0 到 2）
+            z_min, z_max = 0.0, 2.0
+        else:
+            # PPPooling_UAP
+            # 使用 points 的第 3 个通道作为 z 坐标，归一化到 [0, 1]
+            z = points[:, :, 2]  # shape: (B, N)
+            z_min = z.min(dim=1, keepdim=True)[0][0].item()
+            z_max = z.max(dim=1, keepdim=True)[0][0].item()
+            z_range = z_max - z_min + 1e-6
+            z = (z - z_min) / z_range  # shape: (B, N)
+            z_min, z_max = 0.0, 1.0
+
+        all_pooled = []
+        for M in self.bin_num:
+            # 由于 z 已归一化，直接构造均匀分布的 bin 边界
+            edges = torch.linspace(z_min, z_max, steps=M + 1, device=point_clouds.device)
+            # 利用 bucketize 将每个点分配到 [0, M-1] 内的 bin（不需要额外处理首尾）
+            # 注意：这里使用 edges[1:-1] 作为分界，保证边界值归到正确 bin
+            bin_idx = torch.bucketize(z.contiguous(), edges[1:-1], right=False)  # shape: (B, N)
+
+            # 为每个 bin计算 max 和 mean 池化值，利用 scatter_reduce 与 scatter_add 操作：
+            # 构造初始 tensor，形状均为 (B, M, C)
+            pooled_max = torch.full((B, M, C), float('-inf'), device=point_clouds.device, dtype=point_clouds.dtype)
+            pooled_sum = torch.zeros((B, M, C), device=point_clouds.device, dtype=point_clouds.dtype)
+            counts = torch.zeros((B, M, 1), device=point_clouds.device, dtype=point_clouds.dtype)
+
+            # 将 bin_idx 扩展到与 point_clouds 对应的维度 (B, N, C)
+            bin_idx_exp = bin_idx.unsqueeze(-1).expand(-1, -1, C)
+            # max 池化：scatter_reduce 计算每个 bin 内的最大值
+            pooled_max = pooled_max.scatter_reduce(1, bin_idx_exp, point_clouds, reduce='amax', include_self=True)
+            # sum 池化：scatter_add 计算每个 bin 内的和
+            pooled_sum = pooled_sum.scatter_add(1, bin_idx_exp, point_clouds)
+            # 计算每个 bin 的计数
+            counts = counts.scatter_add(1, bin_idx.unsqueeze(-1), torch.ones((B, N, 1), device=point_clouds.device))
+            # 计算 mean 池化
+            pooled_mean = pooled_sum / counts.clamp(min=1)
+            # 这里采用 max 与 mean 的和作为最终池化结果（也可以用 concat）
+            pooled = pooled_max + pooled_mean
+            # 将没有点（max为 -inf）的 bin 置 0
+            pooled[pooled == float('-inf')] = 0
+
+            all_pooled.append(pooled)
+
+        # 将各分辨率下的池化结果在 bin 维度上拼接，并调整为 B x C x M_total
+        output = torch.cat(all_pooled, dim=1)
+        output = rearrange(output, 'B M C -> B C M')
+        return output
+
+
+class NetVLAD(nn.Module):
+    """NetVLAD layer implementation"""
+
+    def __init__(self, num_clusters=64, dim=128, alpha=100.0,
+                 normalize_input=True):
+        """
+        Args:
+            num_clusters : int
+                The number of clusters
+            dim : int
+                Dimension of descriptors
+            alpha : float
+                Parameter of initialization. Larger value is harder assignment.
+            normalize_input : bool
+                If true, descriptor-wise L2 normalization is applied to input.
+        """
+        super(NetVLAD, self).__init__()
+        self.num_clusters = num_clusters
+        self.dim = dim
+        self.alpha = alpha
+        self.normalize_input = normalize_input
+        self.conv = nn.Conv2d(dim, num_clusters, kernel_size=(1, 1), bias=True)
+        self.centroids = nn.Parameter(torch.rand(num_clusters, dim))
+        self._init_params()
+
+    def _init_params(self):
+        self.conv.weight = nn.Parameter(
+            (2.0 * self.alpha * self.centroids).unsqueeze(-1).unsqueeze(-1)
+        )
+        self.conv.bias = nn.Parameter(
+            - self.alpha * self.centroids.norm(dim=1)
+        )
+
+    def forward(self, x, xyz):
+        N, C = x.shape[:2]
+
+        if self.normalize_input:
+            x = F.normalize(x, p=2, dim=1)  # across descriptor dim
+
+        # soft-assignment
+        soft_assign = self.conv(x).view(N, self.num_clusters, -1)
+        soft_assign = F.softmax(soft_assign, dim=1)
+
+        x_flatten = x.view(N, C, -1)
+        
+        # calculate residuals to each clusters
+        residual = x_flatten.expand(self.num_clusters, -1, -1, -1).permute(1, 0, 2, 3) - \
+            self.centroids.expand(x_flatten.size(-1), -1, -1).permute(1, 2, 0).unsqueeze(0)
+        residual *= soft_assign.unsqueeze(2)
+        vlad = residual.sum(dim=-1)
+
+        vlad = F.normalize(vlad, p=2, dim=2)  # intra-normalization
+        vlad = vlad.view(x.size(0), -1)  # flatten  b num c -> b num c
+        vlad = F.normalize(vlad, p=2, dim=1)  # L2 normalize
+        return vlad.unsqueeze(-1)