project/grasp_box/submodules/SAM6D/utils/model_utils.py

import torch
import torch.nn as nn
from torch.nn import functional as F
import open3d as o3d
import trimesh
import copy
import time
import cv2
import logging
import numpy as np
import pytorch3d
from pytorch3d.io import load_objs_as_meshes, load_obj
from pytorch3d.structures import Meshes
from pytorch3d.vis.plotly_vis import AxisArgs, plot_batch_individually, plot_scene
from pytorch3d.vis.texture_vis import texturesuv_image_matplotlib
from pytorch3d.renderer import (
    PerspectiveCameras,
    PointLights, 
    RasterizationSettings, 
    MeshRenderer, 
    MeshRasterizer,  
    SoftPhongShader,
)

from transformers import AutoTokenizer, AutoImageProcessor, AutoModel
from transformers import AutoProcessor, CLIPVisionModelWithProjection
from transformers import CLIPProcessor, CLIPModel
from sklearn.metrics.pairwise import cosine_similarity
import ssl
import os
os.environ['CURL_CA_BUNDLE'] = ''
ssl._create_default_https_context = ssl._create_unverified_context

from pointnet2_utils import (
    gather_operation,
    furthest_point_sample,
)


class LayerNorm2d(nn.Module):
    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
        super().__init__()
        self.weight = nn.Parameter(torch.ones(num_channels))
        self.bias = nn.Parameter(torch.zeros(num_channels))
        self.eps = eps

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        u = x.mean(1, keepdim=True)
        s = (x - u).pow(2).mean(1, keepdim=True)
        x = (x - u) / torch.sqrt(s + self.eps)
        x = self.weight[:, None, None] * x + self.bias[:, None, None]
        return x


def interpolate_pos_embed(model, checkpoint_model):
    if 'pos_embed' in checkpoint_model:
        pos_embed_checkpoint = checkpoint_model['pos_embed']
        embedding_size = pos_embed_checkpoint.shape[-1]
        num_patches = model.patch_embed.num_patches
        num_extra_tokens = model.pos_embed.shape[-2] - num_patches
        # height (== width) for the checkpoint position embedding
        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
        # height (== width) for the new position embedding
        new_size = int(num_patches ** 0.5)
        # class_token and dist_token are kept unchanged
        if orig_size != new_size:
            print("Position interpolate from %dx%d to %dx%d" % (orig_size, orig_size, new_size, new_size))
            extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
            # only the position tokens are interpolated
            pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
            pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
            pos_tokens = torch.nn.functional.interpolate(
                pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
            checkpoint_model['pos_embed'] = new_pos_embed



def sample_pts_feats(pts, feats, npoint=2048, return_index=False):
    '''

        pts: B*N*3

        feats: B*N*C

    '''
    sample_idx = furthest_point_sample(pts, npoint)
    pts = gather_operation(pts.transpose(1,2).contiguous(), sample_idx)
    pts = pts.transpose(1,2).contiguous()
    feats = gather_operation(feats.transpose(1,2).contiguous(), sample_idx)
    feats = feats.transpose(1,2).contiguous()
    if return_index:
        return pts, feats, sample_idx
    else:
        return pts, feats


def get_chosen_pixel_feats(img, choose):
    shape = img.size()
    if len(shape) == 3:
        pass
    elif len(shape) == 4:
        B, C, H, W = shape
        img = img.reshape(B, C, H*W)
    else:
        assert False

    choose = choose.unsqueeze(1).repeat(1, C, 1)
    x = torch.gather(img, 2, choose).contiguous()
    return x.transpose(1,2).contiguous()


def pairwise_distance(

    x: torch.Tensor, y: torch.Tensor, normalized: bool = False, channel_first: bool = False

) -> torch.Tensor:
    r"""Pairwise distance of two (batched) point clouds.



    Args:

        x (Tensor): (*, N, C) or (*, C, N)

        y (Tensor): (*, M, C) or (*, C, M)

        normalized (bool=False): if the points are normalized, we have "x2 + y2 = 1", so "d2 = 2 - 2xy".

        channel_first (bool=False): if True, the points shape is (*, C, N).



    Returns:

        dist: torch.Tensor (*, N, M)

    """
    if channel_first:
        channel_dim = -2
        xy = torch.matmul(x.transpose(-1, -2), y)  # [(*, C, N) -> (*, N, C)] x (*, C, M)
    else:
        channel_dim = -1
        xy = torch.matmul(x, y.transpose(-1, -2))  # (*, N, C) x [(*, M, C) -> (*, C, M)]
    if normalized:
        sq_distances = 2.0 - 2.0 * xy
    else:
        x2 = torch.sum(x ** 2, dim=channel_dim).unsqueeze(-1)  # (*, N, C) or (*, C, N) -> (*, N) -> (*, N, 1)
        y2 = torch.sum(y ** 2, dim=channel_dim).unsqueeze(-2)  # (*, M, C) or (*, C, M) -> (*, M) -> (*, 1, M)
        sq_distances = x2 - 2 * xy + y2
    sq_distances = sq_distances.clamp(min=0.0)
    return sq_distances


def compute_feature_similarity(feat1, feat2, type='cosine', temp=1.0, normalize_feat=True):
    r'''

    Args:

        feat1 (Tensor): (B, N, C)

        feat2 (Tensor): (B, M, C)



    Returns:

        atten_mat (Tensor): (B, N, M)

    '''
    if normalize_feat:
        feat1 = F.normalize(feat1, p=2, dim=2)
        feat2 = F.normalize(feat2, p=2, dim=2)

    if type == 'cosine':
        atten_mat = feat1 @ feat2.transpose(1,2)
    elif type == 'L2':
        atten_mat = torch.sqrt(pairwise_distance(feat1, feat2))
    else:
        assert False

    atten_mat = atten_mat / temp

    return atten_mat

def compute_triangle_normals(pts):
        pts = pts.squeeze(0)
        A = pts[:, 1] - pts[:, 0]  # (6000, 3)
        B = pts[:, 2] - pts[:, 0]  # (6000, 3)
        N = torch.cross(A, B, dim=1)
        normal_magnitude = torch.norm(N, dim=1, keepdim=True)
        return normal_magnitude.unsqueeze(0) 

def aug_pose_noise(gt_r, gt_t,

                std_rots=[15, 10, 5, 1.25, 1],

                max_rot=45,

                sel_std_trans=[0.2, 0.2, 0.2],

                max_trans=0.8):

    B = gt_r.size(0)
    device = gt_r.device

    std_rot = np.random.choice(std_rots)
    angles = torch.normal(mean=0, std=std_rot, size=(B, 3)).to(device=device)
    angles = angles.clamp(min=-max_rot, max=max_rot)
    ones = gt_r.new(B, 1, 1).zero_() + 1
    zeros = gt_r.new(B, 1, 1).zero_()
    a1 = angles[:,0].reshape(B, 1, 1) * np.pi / 180.0
    a1 = torch.cat(
        [torch.cat([torch.cos(a1), -torch.sin(a1), zeros], dim=2),
        torch.cat([torch.sin(a1), torch.cos(a1), zeros], dim=2),
        torch.cat([zeros, zeros, ones], dim=2)], dim=1
    )
    a2 = angles[:,1].reshape(B, 1, 1) * np.pi / 180.0
    a2 = torch.cat(
        [torch.cat([ones, zeros, zeros], dim=2),
        torch.cat([zeros, torch.cos(a2), -torch.sin(a2)], dim=2),
        torch.cat([zeros, torch.sin(a2), torch.cos(a2)], dim=2)], dim=1
    )
    a3 = angles[:,2].reshape(B, 1, 1) * np.pi / 180.0
    a3 = torch.cat(
        [torch.cat([torch.cos(a3), zeros, torch.sin(a3)], dim=2),
        torch.cat([zeros, ones, zeros], dim=2),
        torch.cat([-torch.sin(a3), zeros, torch.cos(a3)], dim=2)], dim=1
    )
    rand_rot = a1 @ a2 @ a3

    rand_trans = torch.normal(
        mean=torch.zeros([B, 3]).to(device),
        std=torch.tensor(sel_std_trans, device=device).view(1, 3),
    )
    rand_trans = torch.clamp(rand_trans, min=-max_trans, max=max_trans)

    rand_rot = gt_r @ rand_rot
    rand_trans = gt_t + rand_trans
    rand_trans[:,2] = torch.clamp(rand_trans[:,2], min=1e-6)

    return rand_rot.detach(), rand_trans.detach()


def compute_coarse_Rt(

    end_points,

    atten,

    pts1,

    pts2,

    depth, 

    radius,

    mask,

    bbox,

    model_pts=None,

    n_proposal1=6000,

    n_proposal2=300,



):

    WSVD = WeightedProcrustes()

    B, N1, _ = pts1.size()
    N2 = pts2.size(1)
    device = pts1.device
    
    # compute soft assignment matrix
    pred_score = torch.softmax(atten, dim=2) * torch.softmax(atten, dim=1)
    pred_label1 = torch.max(pred_score[:,1:,:], dim=2)[1]
    pred_label2 = torch.max(pred_score[:,:,1:], dim=1)[1]
    weights1 = (pred_label1>0).float()
    weights2 = (pred_label2>0).float()

    pred_score = pred_score[:, 1:, 1:].contiguous()
    pred_score = pred_score * weights1.unsqueeze(2) * weights2.unsqueeze(1)
    pred_score = pred_score.reshape(B, N1*N2) ** 1.5

    # sample pose hypothese
    cumsum_weights = torch.cumsum(pred_score, dim=1)
    cumsum_weights /= (cumsum_weights[:, -1].unsqueeze(1).contiguous()+1e-8)
    idx = torch.searchsorted(cumsum_weights, torch.rand(B, n_proposal1*3, device=device))
    idx1, idx2 = idx.div(N2, rounding_mode='floor'), idx % N2
    idx1 = torch.clamp(idx1, max=N1-1).unsqueeze(2).repeat(1,1,3)
    idx2 = torch.clamp(idx2, max=N2-1).unsqueeze(2).repeat(1,1,3)

    p1 = torch.gather(pts1, 1, idx1).reshape(B,n_proposal1,3,3).reshape(B*n_proposal1,3,3)
    p2 = torch.gather(pts2, 1, idx2).reshape(B,n_proposal1,3,3).reshape(B*n_proposal1,3,3)
    
    
    
    
    # # visualize the correspondences
    # rgb_img = end_points['rgb_ori'].cpu().numpy()
    # depth_img = end_points['depth'].cpu().numpy() * 1000
    # intrinsic = end_points['K'].reshape(3,3)
    # rgb = o3d.geometry.Image((rgb_img).astype(np.uint8)) 
    # depth = o3d.geometry.Image((depth_img).astype(np.uint16))
    # rgbd = o3d.geometry.RGBDImage.create_from_color_and_depth(rgb, depth, depth_scale=1000.0)
    # width = rgb_img.shape[1]
    # height = rgb_img.shape[0]
    # cx =  int(intrinsic[0,2])
    # cy =  int(intrinsic[1,2])
    # fx = int(intrinsic[0,0])
    # fy = int(intrinsic[1,1])
    # intri = o3d.camera.PinholeCameraIntrinsic(width=width, height=height, fx=fx, fy=fy, cx=cx, cy=cy) 
    # o3d_points = o3d.geometry.PointCloud.create_from_rgbd_image(rgbd, intrinsic=intri)
    
    # pcd1 = o3d.geometry.PointCloud()
    
    
    
    # o3d.visualization.draw_geometries([o3d_points])
    
    # template_list = [
    # [[1., 0., 0.,], [0., 1., 0.], [0., 0., 1.]], # identity
    # # [[-1., 0., 0.], [0., 1., 0.], [0., 0., 1.]], # mirror along x
    # # [[1., 0., 0.], [0., -1., 0.], [0., 0., 1.]], # mirror along y
    # # [[1., 0., 0.], [0., 1., 0.], [0., 0., -1.]], # mirror along z
    # [[1., 0., 0.], [0., 0., -1.], [0., 1., 0.]], # 90 around x
    # # [[1., 0., 0.], [0., 0., 1.], [0., -1., 0.]], # -90 around x
    # [[0., 0., -1.], [0., 1., 0.], [1., 0., 0.]], # 90 around y
    # # [[0., 0., 1.], [0., 1., 0.], [-1., 0., 0.]], # -90 around y
    # [[0., -1., 0.], [1., 0., 0.], [0., 0., 1.]], # 90 around z
    # # [[0., 1., 0.], [-1., 0., 0.], [0., 0., 1.]] # -90 around z
    # ]
    # template_tensor = torch.tensor(template_list, device='cuda')
    # template_tensor = template_tensor.unsqueeze(0).repeat(n_proposal1,1,1,1).reshape(-1,3,3)
    
    
    pred_rs, pred_ts = WSVD(p2, p1, None)
    pred_rs = pred_rs.reshape(B, n_proposal1, 3, 3)
    pred_ts = pred_ts.reshape(B, n_proposal1, 1, 3)
    
    p1 = p1.reshape(B, n_proposal1, 3, 3)
    p2 = p2.reshape(B, n_proposal1, 3, 3)
    

    # original dis from sam6d
    dis = torch.norm((p1 - pred_ts) @ pred_rs - p2, dim=3).mean(2)   
    
    # add dis from normal vector
    # mesh = end_points['mesh']
    # sample_pts = torch.tensor(trimesh.sample.sample_surface(mesh, 5000)[0], device = 'cuda',dtype=torch.float32)
    # breakpoint()
    # transformed_sample_pts = (pred_rs @ sample_pts.reshape(-1,3).T).T.reshape(-1,3) + pred_ts* (radius + 1e-6)

    # mesh_pts = (p1 - pred_ts) @ pred_rs
    # mesh_normal = compute_triangle_normals(mesh_pts)
    # obs_normal = compute_triangle_normals(p2)
    # eps=1e-6
    # cos_sim = torch.sum(mesh_normal * obs_normal, dim=2).clamp(-1.0 + eps, 1.0 - eps)
    # angle =(cos_sim - cos_sim.min()) / (cos_sim.max() - cos_sim.min() + 1e-8)  # shape (1, 6000)
    # dis_combien = (dis + angle)/2

    ## customize dis 
    mesh = end_points['mesh']
    K =  end_points['K'].reshape(3,3).to(dtype=torch.float64)

    # only project center of CAD model and make sure its depths is not overlimit
    center = torch.tensor(mesh.centroid, device = 'cuda')
    transformed_center = center.reshape(1,1,1,3) + pred_ts* (radius.reshape(-1, 1, 1) + 1e-6)
    Z = transformed_center[..., 2].clamp(min=1e-6)
    
    point_2d = (K @ transformed_center.squeeze().T )
    point_2d[0] = (point_2d[0]/point_2d[2]).int() 
    point_2d[1] = (point_2d[1]/point_2d[2] ).int()
    
    x = point_2d[0].int()
    y = point_2d[1].int()
    
    depth_center = torch.tensor(0.).repeat(n_proposal1).unsqueeze(0).cuda()
    for i in range(n_proposal1):
        try:
            depth_center[0,i] = depth[y[i].item(), x[i].item()]
        except IndexError:
            depth_center[0,i] = -1
    depth_gap = Z.flatten() - depth_center.flatten()
    
    ## given that prediction center deoth should always be larger than gt depth
    # idx_depth =  torch.where(depth_gap > 0)[0]
    
    # if len(idx_depth) == 0:
        
    #     print('all hypothesis are not correct')
    #     pred_R = torch.eye(3).unsqueeze(0).cuda()
    #     pred_t = torch.zeros(1,3).cuda()
        

    #     return pred_R,pred_t,None

    idx_dis = torch.topk(dis, n_proposal2, dim=1, largest=False)[1].sort()[0]
    # idx = torch.tensor(np.intersect1d(idx_dis.cpu().numpy(), idx_depth.cpu().numpy())).cuda()
  

    # if len(idx) == 0:
    #     print('no good selection')
    #     pred_R = torch.eye(3).unsqueeze(0).cuda()
    #     pred_t = torch.zeros(1,3).cuda()
        

    #     return pred_R,pred_t,None
    idx = idx_dis
    idx = idx.squeeze(0)

    pred_rs = torch.gather(pred_rs, 1, idx.reshape(B,idx.shape[0],1,1).repeat(1,1,3,3))
    pred_ts = torch.gather(pred_ts, 1, idx.reshape(B,idx.shape[0],1,1).repeat(1,1,1,3))
    
    p1 = torch.gather(p1, 1, idx.reshape(B,idx.shape[0],1,1).repeat(1,1,3,3))
    p2 = torch.gather(p2, 1, idx.reshape(B,idx.shape[0],1,1).repeat(1,1,3,3))
  
    # # # pose selection
    transformed_pts = (pts1.unsqueeze(1) - pred_ts) @ pred_rs
    transformed_pts = transformed_pts.reshape(B*idx.shape[0], -1, 3)
    if model_pts is None:
        model_pts = pts2
    expand_model_pts = model_pts.unsqueeze(1).repeat(1,idx.shape[0],1,1).reshape(B*idx.shape[0], -1, 3)

    dis = torch.sqrt(pairwise_distance(transformed_pts, expand_model_pts))
    dis = dis.min(2)[0].reshape(B, idx.shape[0], -1)
    
    scores = weights1.unsqueeze(1).sum(2) / ((dis * weights1.unsqueeze(1)).sum(2) +1e-8)

    # add pred-depth vs gt depth comparison score
    
    # cad_points = pts1.reshape(-1,3)
    # # pcd = o3d.geometry.PointCloud()
    # # pcd.points = o3d.utility.Vector3dVector(cad_points)
    # # o3d.visualization.draw_geometries([pcd])
    # scene_pts = pred_ts * (cad_points.reshape(1,1,-1,3)) + pred_ts* (radius + 1e-6)
    
    # rgb_img = end_points['rgb_ori'].to(torch.uint8)
    # depth_img = end_points['depth']* 1000
    # rgb = o3d.geometry.Image(rgb_img.cpu().numpy()) 
    # depth = o3d.geometry.Image(depth_img.cpu().numpy())
     
    # rgbd = o3d.geometry.RGBDImage.create_from_color_and_depth(rgb, depth, depth_scale=1000)
    # width = rgb_img.shape[1]
    # height = rgb_img.shape[0]
    # cx =  int(K[0,2])
    # cy =  int(K[1,2])
    # fx = int(K[0,0])
    # fy = int(K[1,1])
    
    # intri = o3d.camera.PinholeCameraIntrinsic(width=width, height=height, fx=fx, fy=fy, cx=cx, cy=cy) 
    # o3d_points = o3d.geometry.PointCloud.create_from_rgbd_image(rgbd, intrinsic=intri)
    # scene_points = np.array(o3d_points.points)
    
    # cad_points = torch.tensor(trimesh.sample.volume_mesh(mesh, count=500), device='cuda')
    # cad_points_exp = cad_points.unsqueeze(0).unsqueeze(0)  
    # pred_rs_exp = pred_rs.unsqueeze(2)  
    # cad_points_trans = cad_points_exp.transpose(-1, -2)  
    # cad_points = torch.matmul(pred_rs_exp.float(), cad_points_trans.float()) 
    # cad_points = cad_points.squeeze(2).transpose(-1, -2) 
    # cad_points += pred_ts * radius
    
    # cad_points[...,2] = cad_points[...,2] + 1e-6  # avoid divide by 0
    # x = cad_points[..., 0] / cad_points[..., 2]
    # y = cad_points[..., 1] / cad_points[..., 2]

    # U = K[0, 0] * x + K[0, 2]
    # V = K[1, 1] * y + K[1, 2]
    
    # in_bound = (U >= 0) & (U < width) & (V >= 0) & (V < height) 
    
    # cad_points_filtered = cad_points.clone()
    # cad_points_filtered[~in_bound] = np.nan
  
    
    # grid_size = 0.01
    # x, y, z = scene_points[:, 0], scene_points[:, 1], scene_points[:, 2]

    # x_min, y_min = x.min(), y.min()
    # x_idx = ((x - x_min) / grid_size).astype(int)
    # y_idx = ((y - y_min) / grid_size).astype(int)
    # from collections import defaultdict
    # grid = defaultdict(list)
    # for xi, yi, zi in zip(x_idx, y_idx, z):
    #     grid[(xi, yi)].append(zi)

    # averaged_points = []
    

    # for (xi, yi), z_vals in grid.items():
    #     x_center = x_min + (xi + 0.5) * grid_size
    #     y_center = y_min + (yi + 0.5) * grid_size
    #     z_min = np.min(z_vals)
    #     averaged_points.append([x_center, y_center, z_min])

    # averaged_points = np.array(averaged_points)
    
    # pcd = o3d.geometry.PointCloud()
    # pcd.points = o3d.utility.Vector3dVector(averaged_points)
    # pcd.paint_uniform_color([1, 0., 0.]) 
    

    # pcd2 = o3d.geometry.PointCloud()
    # pcd2.points = o3d.utility.Vector3dVector(cad_points_filtered[0,0,:,:].cpu().numpy())
    # pcd2.paint_uniform_color([0, 0.1, 1.]) 
    # o3d.visualization.draw_geometries([pcd,pcd2,o3d_points])
   
    # Step 1: Build the pillar lookup dictionary
    # x_min = averaged_points[:, 0].min()
    # y_min = averaged_points[:, 1].min()

    # pillar_dict = {}
    # for x, y, z in averaged_points:
    #     xi = int((x - x_min) / grid_size)
    #     yi = int((y - y_min) / grid_size)
    #     pillar_dict[(xi, yi)] = z

    # qp_np = cad_points_filtered.squeeze(0).cpu().numpy()  
    # # qp_np = cad_points.squeeze(0).cpu().numpy() 
    # depth_diffs = np.full((qp_np.shape[0], qp_np.shape[1]), np.nan)


    # for i in range(qp_np.shape[0]):
    #     for j in range(qp_np.shape[1]):
    #         x, y, z = qp_np[i, j]
    #         if np.isnan(x) or np.isnan(y) or np.isnan(z):
    #             continue
    #         xi = int((x - x_min) / grid_size)
    #         yi = int((y - y_min) / grid_size)
    #         key = (xi, yi)
    #         if key in pillar_dict:
    #             z_pillar = pillar_dict[key]
    #             depth_diffs[i, j] = z - z_pillar
                

    

    # depth_diffs = torch.tensor(depth_diffs)
    # negative_mask = depth_diffs < 0
    
    # neg_diffs_only = torch.where(negative_mask, depth_diffs, torch.zeros_like(depth_diffs))
    # depth_scores = torch.abs(neg_diffs_only.sum(dim=1))
    # depth_scores = depth_scores / (depth_scores.sum() + 1e-8)
    # depth_scores *= 1000


    # points = torch.tensor(mesh.sample(1000, return_index=False), device='cuda')
    # obs_pts = pred_ts * (points.reshape(1,1,-1,3)) + pred_ts* (radius + 1e-6)
    # obs_depth = obs_pts[..., 2].clamp(min=1e-6)
    
    # obs_img_pts = (K.unsqueeze(0).repeat(len(idx), 1, 1) @ obs_pts.squeeze().permute(0, 2, 1) ).permute(0,2,1)
    # obs_img_pts[:,:,0] = (obs_img_pts[:,:,0]/obs_img_pts[:,:,0]).int() 
    # obs_img_pts[:,:,1] = (obs_img_pts[:,:,1]/obs_img_pts[:,:,2] ).int()
    
    # x = point_2d[0].int()
    # y = point_2d[1].int()
    
    sorted_x, indices = torch.sort(scores,descending=True)
    idx = indices[0,0]
    # print('-----------------',idx,scores[0,idx])
    pred_R = torch.gather(pred_rs, 1, idx.reshape(B,1,1,1).repeat(1,1,3,3)).squeeze(1)
    pred_t = torch.gather(pred_ts, 1, idx.reshape(B,1,1,1).repeat(1,1,1,3)).squeeze(2).squeeze(1)
    pose = np.eye(4)
    pose[:3,:3] = pred_R.cpu().numpy()
    pose[:3,3] = pred_t.cpu().numpy()  *  radius[0].cpu().numpy()
    mesh_o3d = o3d.geometry.TriangleMesh(o3d.utility.Vector3dVector(mesh.vertices),  o3d.utility.Vector3iVector(mesh.faces))
    pcd_obj = mesh_o3d.sample_points_uniformly(number_of_points=1000)
    pcd_obj_trans = copy.deepcopy(pcd_obj).transform(pose)
    pcd_obj_trans.paint_uniform_color([0.1, 0.1, 1]) 
    
    
    if 0:
        scores = scores - depth_scores.cuda()
        sorted_x, indices = torch.sort(scores,descending=True)
        idx = indices[0,1]
        idx = torch.tensor(14)
        # print('-----------------',idx,scores[0,idx])
        pred_R = torch.gather(pred_rs, 1, idx.reshape(B,1,1,1).repeat(1,1,3,3)).squeeze(1)
        pred_t = torch.gather(pred_ts, 1, idx.reshape(B,1,1,1).repeat(1,1,1,3)).squeeze(2).squeeze(1)
        pose = np.eye(4)
        pose[:3,:3] = pred_R.cpu().numpy()
        pose[:3,3] = pred_t.cpu().numpy()  *  radius[0].cpu().numpy()
        mesh_o3d = o3d.geometry.TriangleMesh(o3d.utility.Vector3dVector(mesh.vertices),  o3d.utility.Vector3iVector(mesh.faces))
        pcd_obj2 = mesh_o3d.sample_points_uniformly(number_of_points=1000)
        pcd_obj_trans2 = copy.deepcopy(pcd_obj2).transform(pose)
        pcd_obj_trans2.paint_uniform_color([1, 0.1, 0.1]) 
        o3d.visualization.draw_geometries([pcd_obj_trans,pcd_obj_trans2,o3d_points])
        breakpoint()
    
    #vis the center
    # sphere = o3d.geometry.TriangleMesh.create_sphere(radius=0.02)  
    # sphere.translate()
    # sphere.paint_uniform_color([1, 0, 0])  
    
    
    # pcd2 = o3d.geometry.PointCloud()
    # pcd2.points = o3d.utility.Vector3dVector(qp_np[10])
    # pcd2.paint_uniform_color([0.1, 0.1, 1]) 
    # o3d.visualization.draw_geometries([pcd,pcd2,o3d_points])
    
    
    # approach 2: try project cad points onto 2D img and compare with pillarized gt_depth
    if 0:

        grid_size = 5
        depth_img = end_points['depth']* 1000
        height, width = depth_img.shape
        H_trim = height - height % grid_size
        W_trim = width - width % grid_size
        depth_trimmed = depth_img[:H_trim, :W_trim]

        reshaped = depth_trimmed.reshape(
            H_trim // grid_size, grid_size,
            W_trim // grid_size, grid_size
        )

        reshaped = reshaped.permute(0, 2, 1, 3)
        # min_pillars = reshaped.min(dim=3).values.min(dim=2).values
        n_y, n_x = reshaped.shape[:2]
        pillar_dict = {}

        for yi in range(n_y):
            for xi in range(n_x):
                block = reshaped[yi, xi] 
                min_val = torch.min(block)
                pillar_dict[(xi, yi)] = min_val.item()
        
        # cad_points = torch.tensor(trimesh.sample.volume_mesh(mesh, count=1000), device='cuda')
        cad_points = torch.tensor(mesh.sample(1000), device='cuda')
        cad_points = torch.matmul( pred_rs.squeeze(0).float(), cad_points.T.float()) 
        cad_points = cad_points.transpose(-1, -2) 
        cad_points_trans = cad_points + pred_ts.squeeze(0) * radius
        
        cad_points_trans[...,2] = cad_points_trans[...,2] + 1e-6  # avoid divide by 0
        x = cad_points_trans[..., 0] / cad_points_trans[..., 2]
        y = cad_points_trans[..., 1] / cad_points_trans[..., 2]

        U = K[0, 0] * x + K[0, 2]
        V = K[1, 1] * y + K[1, 2]
        
        
        
        depth_diffs = np.full((cad_points_trans.shape[0], cad_points_trans.shape[1]), np.nan)
        
        
        
        
        for i in range(U.shape[0]):
            for j in range(U.shape[1]):
                x,y = U[i,j], V[i,j]
                z = cad_points_trans[i,j, 2]
                xi = int(x/grid_size)
                yi = int(y/grid_size)
                key = (xi, yi)
                if key in pillar_dict:
                    z_pillar = pillar_dict[key]
                    depth_diffs[i, j] = z - z_pillar
              
        depth_diffs = torch.tensor(depth_diffs)
        negative_mask = depth_diffs < 0
        neg_diffs_only = torch.where(negative_mask, depth_diffs, torch.zeros_like(depth_diffs))
        depth_scores = torch.abs(neg_diffs_only.sum(dim=1))
        depth_scores = depth_scores / (depth_scores.sum() + 1e-8)
        depth_scores *= 1000
        
        scores = weights1.unsqueeze(1).sum(2) / ((dis * weights1.unsqueeze(1)).sum(2) +1e-8)
        scores = scores - (depth_scores).cuda()
        sorted_x, indices = torch.sort(scores,descending=True)
        idx = indices[0,0]
        # print('-----------------',idx,scores[0,idx])
        pred_R = torch.gather(pred_rs, 1, idx.reshape(B,1,1,1).repeat(1,1,3,3)).squeeze(1)
        pred_t = torch.gather(pred_ts, 1, idx.reshape(B,1,1,1).repeat(1,1,1,3)).squeeze(2).squeeze(1)
        pose = np.eye(4)
        pose[:3,:3] = pred_R.cpu().numpy()
        pose[:3,3] = pred_t.cpu().numpy()  *  radius[0].cpu().numpy()
        mesh_o3d = o3d.geometry.TriangleMesh(o3d.utility.Vector3dVector(mesh.vertices),  o3d.utility.Vector3iVector(mesh.faces))
        pcd_obj2 = mesh_o3d.sample_points_uniformly(number_of_points=1000)
        pcd_obj_trans2 = copy.deepcopy(pcd_obj2).transform(pose)
        pcd_obj_trans2.paint_uniform_color([1, 0.1, 0.1]) 
        # o3d.visualization.draw_geometries([pcd_obj_trans,pcd_obj_trans2,o3d_points])
        
        
        # breakpoint()
    
    
    # pcd3 = o3d.geometry.PointCloud()
    # pcd3.points = o3d.utility.Vector3dVector(cad_points_trans[0].cpu().numpy())
    # pcd3.paint_uniform_color([0.1, 0.1, 1]) 
    # o3d.visualization.draw_geometries([pcd,pcd3,o3d_points])
    
    
    # breakpoint()
    # p1 = torch.gather(p1, 1, idx.reshape(B,1,1,1).repeat(1,1,3,3)).squeeze(1)
    # p2 = torch.gather(p2, 1, idx.reshape(B,1,1,1).repeat(1,1,3,3)).squeeze(1)
    # p1 = (p1 - pred_t *(radius.reshape(-1, 1, 1) + 1e-6)) @ pred_R 
    
    # points = p1.cpu().numpy().reshape(-1,3)
    # points = p2.cpu().numpy().reshape(-1,3)
    # p1 = p1.repeat(4,1,1).unsqueeze(0)
    # p2 = p2.repeat(4,1,1).unsqueeze(0)
    
    # pred_R_template = (pred_R @ template_tensor)
    # pcd = torch.tensor(mesh.sample(1000), device='cuda')
    # points_h = torch.hstack([pcd, torch.ones((pcd.shape[0], 1)).cuda()]) 
   
    # pose = torch.eye(4).type(torch.float64).cuda()
    # pose[:3,:3] = pred_R[0]
    # pose[:3,3] = pred_t[0]
    # points_transformed_h = (pose @ points_h.cuda().T).T 
    # points_transformed = points_transformed_h[:, :3]
    
    # mask = end_points['mask_ori']/255
    # rgb_img = end_points['rgb_ori']
    # depth_img = end_points['depth'] * 1000
    
    # rgb_img = rgb_img * mask[:,:,None]
    # depth_img = depth_img * mask
    # rgb = o3d.geometry.Image(rgb_img.cpu().numpy()) 
    # depth = o3d.geometry.Image(depth_img.cpu().numpy())
     
    # rgbd = o3d.geometry.RGBDImage.create_from_color_and_depth(rgb, depth, depth_scale=1000)
    # width = rgb_img.shape[1]
    # height = rgb_img.shape[0]
    # cx =  int(K[0,2])
    # cy =  int(K[1,2])
    # fx = int(K[0,0])
    # fy = int(K[1,1])
    
    # intri = o3d.camera.PinholeCameraIntrinsic(width=width, height=height, fx=fx, fy=fy, cx=cx, cy=cy) 
    # o3d_points = o3d.geometry.PointCloud.create_from_rgbd_image(rgbd, intrinsic=intri)
    # o3d_points.translate(-o3d_points.get_center())
    # breakpoint()  
    # o3d.visualization.draw_geometries([o3d_points])
    # pred_t = pred_t.repeat(4, 1).reshape(1,-1,1,3)
    

    # p1 = (p1 - pred_t) @ pred_R
    # mesh_normal = compute_triangle_normals(p1)
    # obs_normal = compute_triangle_normals(p2)
    # cos_sim = torch.sum(mesh_normal * obs_normal, dim=2).clamp(-1.0 + eps, 1.0 - eps)
    # angle =(cos_sim - cos_sim.min()) / (cos_sim.max() - cos_sim.min() + 1e-8)  # shape (1, 6000)
    
    # add rot error 
    
    
    
    # device = 'cpu'
    # mesh = load_objs_as_meshes(['/workspace/cad_model/box_02/box_02.obj'], device=device)
    # meshes = mesh.extend(10)
    # fx, fy = K[0,0].item(), K[1,1].item()
    # cx, cy = K[0,2].item(), K[1,2].item()
    # h = 480
    # w = 640
    # image_size = torch.tensor([[480, 640]]) 
    
    # RT = torch.eye(4)
    # RT[3,3] = 1
    # RT[:3,:3] = pred_R.reshape(3,3)
    # RT[:3,3] = pred_t * (radius.reshape(-1, 1, 1) + 1e-6)

    # # transfom axis to pytorch3d format
    # Rz = torch.tensor([[-1,0, 0, 0],
    #                 [0, -1, 0, 0],
    #                 [0, 0, 1, 0],
    #                 [0, 0, 0, 1]]).float()

    # RT = torch.matmul(Rz, RT)

    # template_pose_rot = [
    #     torch.tensor([[[1., 0., 0.,], [0., 1., 0.], [0., 0., 1.]]], device = 'cuda'),  # identity
    #     torch.tensor([[[-1., 0., 0.], [0., 1., 0.], [0., 0., 1.]]], device = 'cuda'), # mirror along x
    #     torch.tensor([[[1., 0., 0.], [0., -1., 0.], [0., 0., 1.]]], device = 'cuda'), # mirror along y
    #     torch.tensor([[[1., 0., 0.], [0., 1., 0.], [0., 0., -1.]]], device = 'cuda'), # mirror along z
    #     torch.tensor([[[1., 0., 0.], [0., 0., -1.], [0., 1., 0.]]], device = 'cuda'), # 90 around x
    #     torch.tensor([[[1., 0., 0.], [0., 0., 1.], [0., -1., 0.]]], device = 'cuda'), # -90 around x
    #     torch.tensor([[[0., 0., -1.], [0., 1., 0.], [1., 0., 0.]]], device = 'cuda'), # 90 around y
    #     torch.tensor([[[0., 0., 1.], [0., 1., 0.], [-1., 0., 0.]]], device = 'cuda'), # -90 around y
    #     torch.tensor([[[0., -1., 0.], [1., 0., 0.], [0., 0., 1.]]], device = 'cuda'), # 90 around z
    #     torch.tensor([[[0., 1., 0.], [-1., 0., 0.], [0., 0., 1.]]], device = 'cuda') # -90 around z
    # ]

    # rot_tensor = torch.cat(template_pose_rot, dim=0)   # shape (10,3,3)
    
    # base_rot = RT[:3, :3]   
    # new_rot = base_rot.unsqueeze(0) @ rot_tensor.cpu()           
    
    # RT_batch = RT.unsqueeze(0).repeat(10, 1, 1)
    # RT_batch[:, :3, :3] = new_rot
    
    # R = torch.transpose(RT_batch[:,:3, :3], 1, 2).reshape(-1, 3, 3).detach()
    # T = RT_batch[:,:3,3].reshape(-1,3).detach()
    # f = torch.tensor((fx, fy), dtype=torch.float32).unsqueeze(0)
    # p = torch.tensor((cx, cy), dtype=torch.float32).unsqueeze(0)
   

    # cameras = PerspectiveCameras(
    #             R = R,
    #             T = T,
    #             focal_length=f,
    #             principal_point=p,
    #             image_size=image_size,
    #             in_ndc=False,
    #             device="cpu")
    # raster_settings = RasterizationSettings(
    #         image_size=(h,w), 
    #         blur_radius=0.0, 
    #         faces_per_pixel=10, 
    #     )
    
    # rasterizer = MeshRasterizer(
    #     cameras=cameras, 
    #     raster_settings=raster_settings
    # )
    # lights = PointLights(device=device, location=[[0.0, 0.0, -3.0]])
    # renderer = MeshRenderer(
    #         rasterizer=rasterizer,
    #         shader=SoftPhongShader(
    #             device=device, 
    #             cameras=cameras,
    #             lights=lights
    #         )
    #     )
    
    
    # fragments = rasterizer(meshes)
    # # depths_render = fragments.zbuf.detach().cpu().numpy()
    # depths_render = fragments.zbuf[..., 0].detach().cpu().numpy() 

    
    # images = np.array(renderer(meshes))
    # images = images[:, ..., :3]
    # emb_list = []
    # depth_emb_list = []
    # y1, y2, x1, x2 = end_points['bbox']
    
    # for i in range(images.shape[0]):
    #     img = images[i]
    #     img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
    #     img = (img*255).astype(np.int32)
    #     img = img[y1.int().item():y2.int().item(), x1.int().item():x2.int().item()]
    #     depth_render = depths_render[i][y1.int().item():y2.int().item(), x1.int().item():x2.int().item()]
    #     if len(depth_render.shape) != 3: 
    #         depth_render = np.expand_dims(depth_render, axis=-1) 
    #     depth_render = np.repeat(depth_render*1000, 3, axis=-1).astype(np.uint8)
    #     depth_render_emb = compute_embedding(depth_render)
    #     render_emb = compute_embedding(img)
    #     emb_list.append(render_emb)
    #     depth_emb_list.append(depth_render_emb)
        
    # rgb = end_points['rgb_ori'].cpu().numpy()[y1.int().item():y2.int().item(), x1.int().item():x2.int().item()]
    # depth_crop = depth.unsqueeze(-1).cpu().numpy()[y1.int().item():y2.int().item(), x1.int().item():x2.int().item()]
    # depth_crop = np.repeat(depth_crop*1000, 3, axis=-1).astype(np.uint8)
    # depth_emb = compute_embedding(depth_crop)
    # # get vit embedding of two images (rendering and observation) and compute cosine similarity
    # # rgb =  np.concatenate([rgb, depth_crop], axis=-1)
    # rgb_emb = compute_embedding(rgb)
    # score_sim_rgb = []
    # score_sim_depth = []
    # for idx, emb in enumerate(emb_list):
        
    #     similarity = cosine_similarity(emb.reshape(1,-1).cpu().numpy(), rgb_emb.reshape(1,-1).cpu().numpy())[0][0]
    #     sim_depth = cosine_similarity(depth_emb_list[idx].reshape(1,-1).cpu().numpy(), depth_emb.reshape(1,-1).cpu().numpy())[0][0]
    #     score_sim_rgb.append(similarity)
    #     score_sim_depth.append(sim_depth)
    # # cv2.imwrite('a.png', images[0])

    # final_score = [(a + b) / 2 for a, b in zip(score_sim_rgb, score_sim_depth)]
    # max_index = final_score.index(max(final_score))
    # print('-------------->',max_index)
    # pred_R = pred_R @ template_pose_rot[max_index]
    
    # pred_R = pred_R[idx].unsqueeze(0)
    
    
    return pred_R, pred_t, _
    # return pred_R, pred_t, transformed_center.reshape(6000,3).cpu().numpy()
    # return pred_R, pred_t, (pred_ts_filter* (radius.reshape(-1, 1, 1) + 1e-6)).cpu().numpy().reshape(-1,3)
    # return pred_R, pred_t, (pred_ts_out* (radius.reshape(-1, 1, 1) + 1e-6)).cpu().numpy().reshape(-1,3)
    # return pred_R, pred_t, (pred_ts_filter* (radius.reshape(-1, 1, 1) + 1e-6)).cpu().numpy().reshape(-1,3)
    # return pred_R.unsqueeze(0), pred_t.unsqueeze(0).unsqueeze(0), (pred_t* (radius.reshape(-1, 1, 1) + 1e-6)).cpu().numpy().reshape(-1,3)

def compute_fine_Rt(

    atten,

    pts1,

    pts2,

    radius,

    end_points,

    model_pts=None,

    dis_thres=0.15

):
    if model_pts is None:
        model_pts = pts2

    # compute pose
    WSVD = WeightedProcrustes(weight_thresh=0.0)
    assginment_mat = torch.softmax(atten, dim=2) * torch.softmax(atten, dim=1)
    label1 = torch.max(assginment_mat[:,1:,:], dim=2)[1]
    label2 = torch.max(assginment_mat[:,:,1:], dim=1)[1]

    assginment_mat = assginment_mat[:, 1:, 1:] * (label1>0).float().unsqueeze(2) * (label2>0).float().unsqueeze(1)
    # max_idx = torch.max(assginment_mat, dim=2, keepdim=True)[1]
    # pred_pts = torch.gather(pts2, 1, max_idx.expand_as(pts1))
    normalized_assginment_mat = assginment_mat / (assginment_mat.sum(2, keepdim=True) + 1e-6)
    pred_pts = normalized_assginment_mat @ pts2

    assginment_score = assginment_mat.sum(2)
    pred_R, pred_t = WSVD(pred_pts, pts1, assginment_score)
    # breakpoint()
    # compute score
    pred_pts = (pts1 - pred_t.unsqueeze(1)) @ pred_R
    dis = torch.sqrt(pairwise_distance(pred_pts, model_pts)).min(2)[0]

    mask = (label1>0).float()
    pose_score = (dis < dis_thres).float()
    pose_score = (pose_score * mask).sum(1) / (mask.sum(1) + 1e-8)
    pose_score = pose_score * mask.mean(1)
    
    # add depth filter
    if 0:
        mesh = end_points['mesh']
        K =  end_points['K'].reshape(3,3).to(dtype=torch.float32)

        sample_pts = torch.tensor(trimesh.sample.sample_surface(mesh, 5000)[0], device = 'cuda',dtype=torch.float32)
        transformed_sample_pts = (pred_R @ sample_pts.reshape(-1,3).T).T.reshape(-1,3) + pred_t* (radius + 1e-6)
        Z = transformed_sample_pts[..., 2].clamp(min=1e-6)

        point_2d = (K @ transformed_sample_pts.reshape(-1,3).T )
        point_2d[0] = (point_2d[0]/point_2d[2]).int() 
        point_2d[1] = (point_2d[1]/point_2d[2] ).int()
        
        x = point_2d[0].int()
        y = point_2d[1].int()
        
        
        depth = end_points['depth']
        depth_observation = torch.tensor(0.).repeat(5000).unsqueeze(0).cuda()
        for i in range(5000):
            try:
                depth_observation[0,i] =  depth[y[i].item(), x[i].item()]     
            except IndexError:
                depth_observation[0,i] = -1

        depth_gap = Z- depth_observation
        count = (depth_gap > 0).sum().item()

        # # given that prediction center deoth should always be larger than gt depth
        if count < 2500:
            print('wrong pose from fine network')
            pred_R = torch.eye(3).unsqueeze(0).cuda()
            pred_t = torch.zeros(1,3).cuda()
            return pred_R, pred_t, pose_score, None

    
    # add pose refinement 
    # mesh = end_points['mesh']
    # vertices = torch.tensor(trimesh.bounds.corners(mesh.bounds), device = 'cuda', dtype=torch.float32)
    # K = end_points['K'].reshape(3,3)
    # transformed_vertices = ((pred_R @ vertices.T).squeeze().T) + (pred_t* (radius.reshape(-1, 1, 1) + 1e-6))
    
    # edges_connection = [(0, 1),  (0,3),(0,4), (1,5), (1,2), (2,6), (6,5),(5,1), (4,5), (6,7),(4,7),(3,7)]    
    
    # Z_depth = transformed_vertices[..., 2].clamp(min=1e-6)
    
    # point_2d = (K @ transformed_vertices.squeeze().T )
    # point_2d[0] = (point_2d[0]/point_2d[2]).int() 
    # point_2d[1] = (point_2d[1]/point_2d[2] ).int()
    
    # x = point_2d[0].int().cpu().numpy()
    # y = point_2d[1].int().cpu().numpy()
    
    # depth = end_points['depth'].cpu().numpy()
    # rgb = end_points['rgb_ori'].cpu().numpy().squeeze()
    
    # # get convex hull of object
    # points = np.stack([x, y], axis=-1)

    # n_sample = 50
    # edge_points = []
    # pose_correct = True
    
    # vis = rgb.copy()
    # for start_idx, end_idx in edges_connection:
    #     pt1 = tuple(points[start_idx])
    #     pt2 = tuple(points[end_idx])
    #     cv2.line(vis, pt1, pt2, color=(0, 255, 0), thickness=2)
    # cv2.imwrite('vis.png', vis)
    # transformed_vertices = transformed_vertices.detach().cpu().numpy()
    # for start_idx, end_idx in edges_connection:

    #     edge = np.linspace(points[start_idx], points[end_idx], num=n_sample, dtype=int)
    #     Z_gt = depth[edge[:,1], edge[:,0]]        
    #     Z_object = np.linspace(transformed_vertices[0,start_idx,2], transformed_vertices[0,end_idx,2], num=n_sample, dtype=np.float32)
    #     X_object = np.linspace(transformed_vertices[0,start_idx,0], transformed_vertices[0,end_idx,0], num=n_sample, dtype=np.float32)
    #     Y_object = np.linspace(transformed_vertices[0,start_idx,1], transformed_vertices[0,end_idx,1], num=n_sample, dtype=np.float32)
        
    #     edge_points.append(np.stack([X_object, Y_object, Z_object], axis=-1))
        
    #     depth_gap = Z_gt - Z_object
    #     depth_gap =  np.round(depth_gap, decimals=3)
        
    #     # skip if outside the image
    #     if points[start_idx][0] not in range(0, rgb.shape[1]) or points[start_idx][1] not in range(0, rgb.shape[0]):
    #         continue
        
    #     if points[end_idx][0] not in range(0, rgb.shape[1]) or points[end_idx][1] not in range(0, rgb.shape[0]):
    #         continue
        
    #     # visualize the edge 
    #     vis = rgb.copy()
    #     cv2.line(vis, tuple(points[start_idx]), tuple(points[end_idx]), (0, 0, 255)  , 3)
    #     cv2.imwrite('vis.png', vis)
     
    #     # print(depth_gap)
    #     if len(depth_gap[depth_gap < -0.02])>=int(0.5*n_sample):
    #         # print('edge is occluded')
    #         continue
        
    #     not_occluded_depth = depth_gap[depth_gap >= -0.05]
    #     # set a threshold for depth gap. if larger than 0.05 (5cm), pose is wrong
    #     # print('num of hover points: ', len(not_occluded_depth[not_occluded_depth> 0.05]))
    #     # if len(not_occluded_depth[not_occluded_depth> 0.05]) >= int(0.2*n_sample):
    #     #     print('-------------------> too much hover points!!!<------------------')
    #     #     pose_correct = False
    #     #     break
    #     edge = edge[np.where(not_occluded_depth < 0.05)[0]]
    #     #now check whether the projected edge is edge or not on the rgbd image. if not, pose is wrong
    #     # gray = cv2.cvtColor(depth, cv2.COLOR_BGR2GRAY).astype(np.uint8)
    #     edges = cv2.Canny((depth*1000).astype(np.uint8), threshold1=50, threshold2=150)
    #     x1,y1 = edge[0]
    #     x2,y2 = edge[-1]
    
    #     edges_mask = np.zeros_like(depth)
    #     # edges = np.stack((edges,)*3, axis=-1)
    #     length_projected_edge =int( (abs(x2-x1)**2 + abs(y2-y1)**2)**0.5)
    #     # cv2.line(edges_mask, tuple(points[i]), tuple(points[i+1]), (1, 1, 1)  , 8)

    #     masked_edges = edges * edges_mask
    #     valid = np.sum( masked_edges/255)/length_projected_edge
    #     print('valid:', valid)
    #     if valid < 1:
    #         print('edge is not edge on the rgbd image!!!')
    #         pose_correct = False
    # print('pose is correct:',pose_correct)
    # edge_points = np.array(edge_points).reshape(-1,3)
    
    return pred_R, pred_t, pose_score, None

def compute_embedding(image):
    
    if 1: # using CLIP
        clip_model = "openai/clip-vit-base-patch32"
        vision_model = CLIPVisionModelWithProjection.from_pretrained(clip_model)
        model = CLIPModel.from_pretrained ("openai/clip-vit-base-patch32")
        processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
        inputs = processor(images=[image], return_tensors="pt", padding=True)
        inputs = inputs.to('cpu')
        image_outputs = vision_model(**inputs)
        img_feats = image_outputs.image_embeds.view(1, -1)
        img_feats = img_feats / img_feats.norm(p=2, dim=-1, keepdim=True)
        
        return img_feats.detach()
    if 0: # using DINOv2
        processor = AutoImageProcessor.from_pretrained('facebook/dinov2-small')
        model = AutoModel.from_pretrained('facebook/dinov2-small').to('cuda')
        with torch.no_grad():
            inputs = processor(images= image, return_tensors = "pt").to('cuda')
            outputs = model(**inputs)
            embedding = outputs.last_hidden_state.mean(dim=1)
            return embedding
        
    

def backproject_points(xy_points, depth_image,K):
    
    K_inv = np.linalg.inv(K)
    points_3d = []
    
    for (u, v) in xy_points:
    
        u = int(round(u))
        v = int(round(v))
        

        z = depth_image[v, u]
        pixel_homog = np.array([u, v, 1.0])
        pt_3d = z * (K_inv @ pixel_homog)
        points_3d.append(pt_3d)
    
    return np.array(points_3d)

def weighted_procrustes(

    src_points,

    ref_points,

    weights=None,

    weight_thresh=0.0,

    eps=1e-5,

    return_transform=False,

    src_centroid = None,

    ref_centroid = None,

):
    r"""Compute rigid transformation from `src_points` to `ref_points` using weighted SVD.



    Modified from [PointDSC](https://github.com/XuyangBai/PointDSC/blob/master/models/common.py).



    Args:

        src_points: torch.Tensor (B, N, 3) or (N, 3)

        ref_points: torch.Tensor (B, N, 3) or (N, 3)

        weights: torch.Tensor (B, N) or (N,) (default: None)

        weight_thresh: float (default: 0.)

        eps: float (default: 1e-5)

        return_transform: bool (default: False)



    Returns:

        R: torch.Tensor (B, 3, 3) or (3, 3)

        t: torch.Tensor (B, 3) or (3,)

        transform: torch.Tensor (B, 4, 4) or (4, 4)

    """
    if src_points.ndim == 2:
        src_points = src_points.unsqueeze(0)
        ref_points = ref_points.unsqueeze(0)
        if weights is not None:
            weights = weights.unsqueeze(0)
        squeeze_first = True
    else:
        squeeze_first = False

    batch_size = src_points.shape[0]
    if weights is None:
        weights = torch.ones_like(src_points[:, :, 0])
    weights = torch.where(torch.lt(weights, weight_thresh), torch.zeros_like(weights), weights)
    weights = weights / (torch.sum(weights, dim=1, keepdim=True) + eps)
    weights = weights.unsqueeze(2)  # (B, N, 1)

    if src_centroid is None:
        src_centroid = torch.sum(src_points * weights, dim=1, keepdim=True)  # (B, 1, 3)
    elif len(src_centroid.size()) == 2:
        src_centroid = src_centroid.unsqueeze(1)
    src_points_centered = src_points - src_centroid  # (B, N, 3)

    if ref_centroid is None:
        ref_centroid = torch.sum(ref_points * weights, dim=1, keepdim=True)  # (B, 1, 3)
    elif len(ref_centroid.size()) == 2:
        ref_centroid = ref_centroid.unsqueeze(1)
    ref_points_centered = ref_points - ref_centroid  # (B, N, 3)

    H = src_points_centered.permute(0, 2, 1) @ (weights * ref_points_centered)
    U, _, V = torch.svd(H)
    Ut, V = U.transpose(1, 2), V
    eye = torch.eye(3).unsqueeze(0).repeat(batch_size, 1, 1).to(src_points.device)
    eye[:, -1, -1] = torch.sign(torch.det(V @ Ut))
    R = V @ eye @ Ut

    t = ref_centroid.permute(0, 2, 1) - R @ src_centroid.permute(0, 2, 1)
    t = t.squeeze(2)

    if return_transform:
        transform = torch.eye(4).unsqueeze(0).repeat(batch_size, 1, 1).cuda()
        transform[:, :3, :3] = R
        transform[:, :3, 3] = t
        if squeeze_first:
            transform = transform.squeeze(0)
        return transform
    else:
        if squeeze_first:
            R = R.squeeze(0)
            t = t.squeeze(0)
        return R, t


class WeightedProcrustes(nn.Module):
    def __init__(self, weight_thresh=0.5, eps=1e-5, return_transform=False):
        super(WeightedProcrustes, self).__init__()
        self.weight_thresh = weight_thresh
        self.eps = eps
        self.return_transform = return_transform

    def forward(self, src_points, tgt_points, weights=None,src_centroid = None,ref_centroid = None):
        return weighted_procrustes(
            src_points,
            tgt_points,
            weights=weights,
            weight_thresh=self.weight_thresh,
            eps=self.eps,
            return_transform=self.return_transform,
            src_centroid=src_centroid,
            ref_centroid=ref_centroid
        )