# # Copyright (C) 2023, Inria # GRAPHDECO research group, https://team.inria.fr/graphdeco # All rights reserved. # # This software is free for non-commercial, research and evaluation use # under the terms of the LICENSE.md file. # # For inquiries contact george.drettakis@inria.fr # import os import numpy as np import torch from plyfile import PlyData, PlyElement from pytorch3d.transforms import quaternion_to_matrix from simple_knn._C import distCUDA2 from torch import nn from field_construction.scene.per_point_adam import PerPointAdam from field_construction.utils.general_utils import (build_rotation, build_scaling, build_scaling_rotation, get_expon_lr_func, inverse_sigmoid, strip_symmetric) from field_construction.utils.graphics_utils import BasicPointCloud from field_construction.utils.pose_utils import get_tensor_from_camera from field_construction.utils.sh_utils import RGB2SH from field_construction.utils.system_utils import mkdir_p def dilate(bin_img, ksize=5): pad = (ksize - 1) // 2 bin_img = torch.nn.functional.pad(bin_img, pad=[pad, pad, pad, pad], mode='reflect') out = torch.nn.functional.max_pool2d(bin_img, kernel_size=ksize, stride=1, padding=0) return out def erode(bin_img, ksize=5): out = 1 - dilate(1 - bin_img, ksize) return out class GaussianModel: def setup_functions(self): def build_covariance_from_scaling_rotation(scaling, scaling_modifier, rotation): L = build_scaling_rotation(scaling_modifier * scaling, rotation) actual_covariance = L @ L.transpose(1, 2) symm = strip_symmetric(actual_covariance) return symm self.scaling_activation = torch.exp self.scaling_inverse_activation = torch.log self.covariance_activation = build_covariance_from_scaling_rotation self.opacity_activation = torch.sigmoid self.inverse_opacity_activation = inverse_sigmoid self.rotation_activation = torch.nn.functional.normalize def __init__(self, sh_degree : int): self.active_sh_degree = 0 self.max_sh_degree = sh_degree self._xyz = torch.empty(0) self._knn_f = torch.empty(0) self._features_dc = torch.empty(0) self._features_rest = torch.empty(0) self._scaling = torch.empty(0) self._rotation = torch.empty(0) self._opacity = torch.empty(0) self._language_feature = torch.empty(0) self._instance_feature=torch.empty(0) self.max_radii2D = torch.empty(0) self.max_weight = torch.empty(0) self.xyz_gradient_accum = torch.empty(0) self.xyz_gradient_accum_abs = torch.empty(0) self.denom = torch.empty(0) self.denom_abs = torch.empty(0) self.optimizer = None self.cam_optimizer = None self.percent_dense = 0 self.spatial_lr_scale = 0 self.knn_dists = None self.knn_idx = None self.setup_functions() self.use_app = False def capture(self, include_feature=False): if include_feature: return ( self.active_sh_degree, self._xyz, self._knn_f, self._features_dc, self._features_rest, self._scaling, self._rotation, self._opacity, self._language_feature, self._instance_feature, self.max_radii2D, self.max_weight, self.xyz_gradient_accum, self.xyz_gradient_accum_abs, self.denom, self.denom_abs, self.optimizer.state_dict(), self.cam_optimizer.state_dict(), self.spatial_lr_scale, self.P ) else: return ( self.active_sh_degree, self._xyz, self._knn_f, self._features_dc, self._features_rest, self._scaling, self._rotation, self._opacity, self.max_radii2D, self.max_weight, self.xyz_gradient_accum, self.xyz_gradient_accum_abs, self.denom, self.denom_abs, self.optimizer.state_dict(), self.cam_optimizer.state_dict(), self.spatial_lr_scale, self.P ) def restore(self, model_args, training_args, mode='train'): # Ckpt with training feature (20 arguments) if len(model_args) == 20: (self.active_sh_degree, self._xyz, self._knn_f, self._features_dc, self._features_rest, self._scaling, self._rotation, self._opacity, self._language_feature, # Added training feature: language feature self._instance_feature, # Added training feature: instance feature self.max_radii2D, self.max_weight, xyz_gradient_accum, xyz_gradient_accum_abs, denom, denom_abs, opt_dict, cam_opt_dict, self.spatial_lr_scale, self.P ) = model_args # Ckpt without training feature (18 arguments) elif len(model_args) == 18: (self.active_sh_degree, self._xyz, self._knn_f, self._features_dc, self._features_rest, self._scaling, self._rotation, self._opacity, self.max_radii2D, self.max_weight, xyz_gradient_accum, xyz_gradient_accum_abs, denom, denom_abs, opt_dict, cam_opt_dict, self.spatial_lr_scale, self.P ) = model_args if mode == 'train': if isinstance(self.optimizer, PerPointAdam): self.training_setup_pp(training_args) else: self.training_setup(training_args) self.xyz_gradient_accum = xyz_gradient_accum self.xyz_gradient_accum_abs = xyz_gradient_accum_abs self.denom = denom self.denom_abs = denom_abs self.optimizer.load_state_dict(opt_dict) self.cam_optimizer.load_state_dict(cam_opt_dict) @property def get_scaling(self): return self.scaling_activation(self._scaling) @property def get_rotation(self): return self.rotation_activation(self._rotation) @property def get_xyz(self): return self._xyz @property def get_features(self): features_dc = self._features_dc features_rest = self._features_rest return torch.cat((features_dc, features_rest), dim=1) @property def get_opacity(self): return self.opacity_activation(self._opacity) @property def get_language_feature(self): return self._language_feature @property def get_instance_feature(self): return self._instance_feature def get_smallest_axis(self, return_idx=False): rotation_matrices = self.get_rotation_matrix() smallest_axis_idx = self.get_scaling.min(dim=-1)[1][..., None, None].expand(-1, 3, -1) smallest_axis = rotation_matrices.gather(2, smallest_axis_idx) if return_idx: return smallest_axis.squeeze(dim=2), smallest_axis_idx[..., 0, 0] return smallest_axis.squeeze(dim=2) def get_normal(self, view_cam): normal_global = self.get_smallest_axis() gaussian_to_cam_global = view_cam.camera_center - self._xyz neg_mask = (normal_global * gaussian_to_cam_global).sum(-1) < 0.0 normal_global[neg_mask] = -normal_global[neg_mask] return normal_global def init_RT_seq(self, cam_list): poses =[] index_mapping = {} for cam_idx, cam in enumerate(cam_list[1.0]): p = get_tensor_from_camera(cam.world_view_transform.transpose(0, 1)) # R T -> quat t poses.append(p) index_mapping[cam.uid] = cam_idx poses = torch.stack(poses) self.index_mapping = index_mapping self.P = poses.cuda().requires_grad_(True) def get_RT(self, idx): pose = self.P[idx] return pose def get_RT_test(self, idx): pose = self.test_P[idx] return pose def get_rotation_matrix(self): return quaternion_to_matrix(self.get_rotation) def get_covariance(self, scaling_modifier = 1): return self.covariance_activation(self.get_scaling, scaling_modifier, self._rotation) def oneupSHdegree(self): if self.active_sh_degree < self.max_sh_degree: self.active_sh_degree += 1 def create_from_pcd(self, pcd : BasicPointCloud, spatial_lr_scale : float): self.spatial_lr_scale = spatial_lr_scale fused_point_cloud = torch.tensor(np.asarray(pcd.points)).float().cuda() fused_color = RGB2SH(torch.tensor(np.asarray(pcd.colors)).float().cuda()) features = torch.zeros((fused_color.shape[0], 3, (self.max_sh_degree + 1) ** 2)).float().cuda() features[:, :3, 0 ] = fused_color features[:, 3:, 1:] = 0.0 print("Number of points at initialisation : ", fused_point_cloud.shape[0]) dist = torch.sqrt(torch.clamp_min(distCUDA2(torch.from_numpy(np.asarray(pcd.points)).float().cuda()), 0.0000001)) # print(f"new scale {torch.quantile(dist, 0.1)}") scales = torch.log(dist)[...,None].repeat(1, 3) rots = torch.zeros((fused_point_cloud.shape[0], 4), device="cuda") rots[:, 0] = 1 opacities = inverse_sigmoid(0.1 * torch.ones((fused_point_cloud.shape[0], 1), dtype=torch.float, device="cuda")) knn_f = torch.randn((fused_point_cloud.shape[0], 6)).float().cuda() self._xyz = nn.Parameter(fused_point_cloud.requires_grad_(True)) self._knn_f = nn.Parameter(knn_f.requires_grad_(True)) self._features_dc = nn.Parameter(features[:,:,0:1].transpose(1, 2).contiguous().requires_grad_(True)) self._features_rest = nn.Parameter(features[:,:,1:].transpose(1, 2).contiguous().requires_grad_(True)) self._scaling = nn.Parameter(scales.requires_grad_(True)) self._rotation = nn.Parameter(rots.requires_grad_(True)) self._opacity = nn.Parameter(opacities.requires_grad_(True)) self.max_radii2D = torch.zeros((self.get_xyz.shape[0]), device="cuda") self.max_weight = torch.zeros((self.get_xyz.shape[0]), device="cuda") language_feature = torch.zeros((fused_point_cloud.shape[0], 3), device="cuda") self._language_feature = nn.Parameter(language_feature.requires_grad_(True)).requires_grad_(True) # dont train feature at first # NOTE for instance distinguish instance_feature = torch.zeros((fused_point_cloud.shape[0], 3), device="cuda") self._instance_feature = nn.Parameter(instance_feature.requires_grad_(False)).requires_grad_(False) # just train feature at last def training_setup(self, training_args, device): self.percent_dense = training_args.percent_dense self.xyz_gradient_accum = torch.zeros((self.get_xyz.shape[0], 1), device=device) self.xyz_gradient_accum_abs = torch.zeros((self.get_xyz.shape[0], 1), device=device) self.denom = torch.zeros((self.get_xyz.shape[0], 1), device=device) self.denom_abs = torch.zeros((self.get_xyz.shape[0], 1), device=device) self.abs_split_radii2D_threshold = training_args.abs_split_radii2D_threshold self.max_abs_split_points = training_args.max_abs_split_points self.max_all_points = training_args.max_all_points l = [ {'params': [self._xyz], 'lr': training_args.position_lr_init * self.spatial_lr_scale, "name": "xyz"}, {'params': [self._knn_f], 'lr': 0.01, "name": "knn_f"}, {'params': [self._features_dc], 'lr': training_args.feature_lr, "name": "f_dc"}, {'params': [self._features_rest], 'lr': training_args.feature_lr / 20.0, "name": "f_rest"}, {'params': [self._opacity], 'lr': training_args.opacity_lr, "name": "opacity"}, {'params': [self._scaling], 'lr': training_args.scaling_lr, "name": "scaling"}, {'params': [self._rotation], 'lr': training_args.rotation_lr, "name": "rotation"}, {'params': [self._language_feature], 'lr': training_args.language_feature_lr, "name": "language_feature"}, # semantic {'params': [self._instance_feature], 'lr': training_args.language_feature_lr, "name": "instance_feature"}, # instance ] l_cam = [{'params': [self.P],'lr': training_args.rotation_lr*0.1, "name": "pose"},] # l += l_cam self.optimizer = torch.optim.Adam(l, lr=0.0, eps=1e-15) self.cam_optimizer = torch.optim.Adam(l_cam, lr=0.0, eps=1e-15) self.xyz_scheduler_args = get_expon_lr_func(lr_init=training_args.position_lr_init*self.spatial_lr_scale, lr_final=training_args.position_lr_final*self.spatial_lr_scale, lr_delay_mult=training_args.position_lr_delay_mult, max_steps=training_args.position_lr_max_steps) self.cam_scheduler_args = get_expon_lr_func( lr_init=training_args.rotation_lr*0.1, lr_final=training_args.rotation_lr*0.001, lr_delay_mult=training_args.position_lr_delay_mult, max_steps=training_args.iterations) # per-point optimizer def training_setup_pp(self, training_args, confidence_lr=None, device="cuda"): self.percent_dense = training_args.percent_dense self.xyz_gradient_accum = torch.zeros((self.get_xyz.shape[0], 1), device=device) self.xyz_gradient_accum_abs = torch.zeros((self.get_xyz.shape[0], 1), device=device) self.denom = torch.zeros((self.get_xyz.shape[0], 1), device=device) self.denom_abs = torch.zeros((self.get_xyz.shape[0], 1), device=device) self.abs_split_radii2D_threshold = training_args.abs_split_radii2D_threshold self.max_abs_split_points = training_args.max_abs_split_points self.max_all_points = training_args.max_all_points self.per_point_lr = confidence_lr l = [ {'params': [self._xyz], 'per_point_lr': self.per_point_lr, 'lr': training_args.position_lr_init * self.spatial_lr_scale, "name": "xyz"}, {'params': [self._knn_f], 'lr': 0.01, "name": "knn_f"}, {'params': [self._features_dc], 'lr': training_args.feature_lr, "name": "f_dc"}, {'params': [self._features_rest], 'lr': training_args.feature_lr / 20.0, "name": "f_rest"}, {'params': [self._opacity], 'lr': training_args.opacity_lr, "name": "opacity"}, {'params': [self._scaling], 'lr': training_args.scaling_lr, "name": "scaling"}, {'params': [self._rotation], 'lr': training_args.rotation_lr, "name": "rotation"}, {'params': [self._language_feature], 'lr': training_args.language_feature_lr, "name": "language_feature"}, # semantic {'params': [self._instance_feature], 'lr': training_args.language_feature_lr, "name": "instance_feature"}, # instance ] l_cam = [{'params': [self.P],'lr': training_args.rotation_lr*0.1, "name": "pose"},] # l += l_cam self.optimizer = PerPointAdam(l, lr=0, betas=(0.9, 0.999), eps=1e-15, weight_decay=0.0) self.cam_optimizer = torch.optim.Adam(l_cam, lr=0.0, eps=1e-15) self.xyz_scheduler_args = get_expon_lr_func(lr_init=training_args.position_lr_init*self.spatial_lr_scale, lr_final=training_args.position_lr_final*self.spatial_lr_scale, lr_delay_mult=training_args.position_lr_delay_mult, max_steps=training_args.position_lr_max_steps) self.cam_scheduler_args = get_expon_lr_func( lr_init=training_args.rotation_lr*0.1, lr_final=training_args.rotation_lr*0.001, lr_delay_mult=training_args.position_lr_delay_mult, max_steps=training_args.iterations) def clip_grad(self, norm=1.0): for group in self.optimizer.param_groups: torch.nn.utils.clip_grad_norm_(group["params"][0], norm) def update_learning_rate(self, iteration): ''' Learning rate scheduling per step ''' for param_group in self.cam_optimizer.param_groups: if param_group["name"] == "pose": lr = self.cam_scheduler_args(iteration) param_group['lr'] = lr for param_group in self.optimizer.param_groups: if param_group["name"] == "xyz": lr = self.xyz_scheduler_args(iteration) param_group['lr'] = lr def construct_list_of_attributes(self, include_feature=False): l = ['x', 'y', 'z', 'nx', 'ny', 'nz'] # All channels except the 3 DC for i in range(self._features_dc.shape[1]*self._features_dc.shape[2]): l.append('f_dc_{}'.format(i)) for i in range(self._features_rest.shape[1]*self._features_rest.shape[2]): l.append('f_rest_{}'.format(i)) l.append('opacity') for i in range(self._scaling.shape[1]): l.append('scale_{}'.format(i)) for i in range(self._rotation.shape[1]): l.append('rot_{}'.format(i)) if include_feature: for i in range(self._language_feature.shape[1]): l.append('language_feature_{}'.format(i)) for i in range(self._instance_feature.shape[1]): l.append('instance_feature_{}'.format(i)) return l def save_ply(self, path, mask=None, include_feature=False): mkdir_p(os.path.dirname(path)) xyz = self._xyz.detach().cpu().numpy() normals = np.zeros_like(xyz) f_dc = self._features_dc.detach().transpose(1, 2).flatten(start_dim=1).contiguous().cpu().numpy() f_rest = self._features_rest.detach().transpose(1, 2).flatten(start_dim=1).contiguous().cpu().numpy() opacities = self._opacity.detach().cpu().numpy() scale = self._scaling.detach().cpu().numpy() rotation = self._rotation.detach().cpu().numpy() language_feature = self._language_feature.detach().cpu().numpy() instance_feature = self._instance_feature.detach().cpu().numpy() dtype_full = [(attribute, 'f4') for attribute in self.construct_list_of_attributes(include_feature)] elements = np.empty(xyz.shape[0], dtype=dtype_full) if include_feature: attributes = np.concatenate((xyz, normals, f_dc, f_rest, opacities, scale, rotation, language_feature, instance_feature), axis=1) else: attributes = np.concatenate((xyz, normals, f_dc, f_rest, opacities, scale, rotation), axis=1) elements[:] = list(map(tuple, attributes)) el = PlyElement.describe(elements, 'vertex') PlyData([el]).write(path) def reset_opacity(self): opacities_new = inverse_sigmoid(torch.min(self.get_opacity, torch.ones_like(self.get_opacity)*0.01)) optimizable_tensors = self.replace_tensor_to_optimizer(opacities_new, "opacity") self._opacity = optimizable_tensors["opacity"] def load_ply(self, path): plydata = PlyData.read(path) xyz = np.stack((np.asarray(plydata.elements[0]["x"]), np.asarray(plydata.elements[0]["y"]), np.asarray(plydata.elements[0]["z"])), axis=1) opacities = np.asarray(plydata.elements[0]["opacity"])[..., np.newaxis] features_dc = np.zeros((xyz.shape[0], 3, 1)) features_dc[:, 0, 0] = np.asarray(plydata.elements[0]["f_dc_0"]) features_dc[:, 1, 0] = np.asarray(plydata.elements[0]["f_dc_1"]) features_dc[:, 2, 0] = np.asarray(plydata.elements[0]["f_dc_2"]) extra_f_names = [p.name for p in plydata.elements[0].properties if p.name.startswith("f_rest_")] extra_f_names = sorted(extra_f_names, key = lambda x: int(x.split('_')[-1])) assert len(extra_f_names)==3*(self.max_sh_degree + 1) ** 2 - 3 features_extra = np.zeros((xyz.shape[0], len(extra_f_names))) for idx, attr_name in enumerate(extra_f_names): features_extra[:, idx] = np.asarray(plydata.elements[0][attr_name]) # Reshape (P,F*SH_coeffs) to (P, F, SH_coeffs except DC) features_extra = features_extra.reshape((features_extra.shape[0], 3, (self.max_sh_degree + 1) ** 2 - 1)) scale_names = [p.name for p in plydata.elements[0].properties if p.name.startswith("scale_")] scale_names = sorted(scale_names, key = lambda x: int(x.split('_')[-1])) scales = np.zeros((xyz.shape[0], len(scale_names))) for idx, attr_name in enumerate(scale_names): scales[:, idx] = np.asarray(plydata.elements[0][attr_name]) rot_names = [p.name for p in plydata.elements[0].properties if p.name.startswith("rot")] rot_names = sorted(rot_names, key = lambda x: int(x.split('_')[-1])) rots = np.zeros((xyz.shape[0], len(rot_names))) for idx, attr_name in enumerate(rot_names): rots[:, idx] = np.asarray(plydata.elements[0][attr_name]) language_feature_names = [p.name for p in plydata.elements[0].properties if p.name.startswith("language_feature")] language_feature_names = sorted(language_feature_names, key = lambda x: int(x.split('_')[-1])) language_feature = np.zeros((xyz.shape[0], len(language_feature_names))) for idx, attr_name in enumerate(language_feature_names): language_feature[:, idx] = np.asarray(plydata.elements[0][attr_name]) # NOTE instance instance_feature_names = [p.name for p in plydata.elements[0].properties if p.name.startswith("instance_feature")] instance_feature_names = sorted(instance_feature_names, key = lambda x: int(x.split('_')[-1])) instance_feature = np.zeros((xyz.shape[0], len(instance_feature_names))) for idx, attr_name in enumerate(instance_feature_names): instance_feature[:, idx] = np.asarray(plydata.elements[0][attr_name]) self._xyz = nn.Parameter(torch.tensor(xyz, dtype=torch.float, device="cuda").requires_grad_(True)) self._features_dc = nn.Parameter(torch.tensor(features_dc, dtype=torch.float, device="cuda").transpose(1, 2).contiguous().requires_grad_(True)) self._features_rest = nn.Parameter(torch.tensor(features_extra, dtype=torch.float, device="cuda").transpose(1, 2).contiguous().requires_grad_(True)) self._opacity = nn.Parameter(torch.tensor(opacities, dtype=torch.float, device="cuda").requires_grad_(True)) self._scaling = nn.Parameter(torch.tensor(scales, dtype=torch.float, device="cuda").requires_grad_(True)) self._rotation = nn.Parameter(torch.tensor(rots, dtype=torch.float, device="cuda").requires_grad_(True)) self._language_feature = nn.Parameter(torch.tensor(language_feature, dtype=torch.float, device="cuda").requires_grad_(False)) self._instance_feature = nn.Parameter(torch.tensor(instance_feature, dtype=torch.float, device="cuda").requires_grad_(False)) self.active_sh_degree = self.max_sh_degree def replace_tensor_to_optimizer(self, tensor, name): optimizable_tensors = {} for group in self.optimizer.param_groups: if group["name"] == name: stored_state = self.optimizer.state.get(group['params'][0], None) stored_state["exp_avg"] = torch.zeros_like(tensor) stored_state["exp_avg_sq"] = torch.zeros_like(tensor) del self.optimizer.state[group['params'][0]] group["params"][0] = nn.Parameter(tensor.requires_grad_(True)) self.optimizer.state[group['params'][0]] = stored_state optimizable_tensors[group["name"]] = group["params"][0] return optimizable_tensors def _prune_optimizer(self, mask): optimizable_tensors = {} for group in self.optimizer.param_groups: stored_state = self.optimizer.state.get(group['params'][0], None) if stored_state is not None: stored_state["exp_avg"] = stored_state["exp_avg"][mask] stored_state["exp_avg_sq"] = stored_state["exp_avg_sq"][mask] del self.optimizer.state[group['params'][0]] group["params"][0] = nn.Parameter((group["params"][0][mask].requires_grad_(True))) self.optimizer.state[group['params'][0]] = stored_state optimizable_tensors[group["name"]] = group["params"][0] else: group["params"][0] = nn.Parameter(group["params"][0][mask].requires_grad_(True)) optimizable_tensors[group["name"]] = group["params"][0] return optimizable_tensors def prune_points(self, mask): valid_points_mask = ~mask optimizable_tensors = self._prune_optimizer(valid_points_mask) self._xyz = optimizable_tensors["xyz"] self._knn_f = optimizable_tensors["knn_f"] self._features_dc = optimizable_tensors["f_dc"] self._features_rest = optimizable_tensors["f_rest"] self._opacity = optimizable_tensors["opacity"] self._scaling = optimizable_tensors["scaling"] self._rotation = optimizable_tensors["rotation"] self._language_feature = optimizable_tensors["language_feature"] self._instance_feature = optimizable_tensors["instance_feature"] self.xyz_gradient_accum = self.xyz_gradient_accum[valid_points_mask] self.xyz_gradient_accum_abs = self.xyz_gradient_accum_abs[valid_points_mask] self.denom = self.denom[valid_points_mask] self.denom_abs = self.denom_abs[valid_points_mask] self.max_radii2D = self.max_radii2D[valid_points_mask] self.max_weight = self.max_weight[valid_points_mask] def cat_tensors_to_optimizer(self, tensors_dict): optimizable_tensors = {} for group in self.optimizer.param_groups: assert len(group["params"]) == 1 extension_tensor = tensors_dict[group["name"]] stored_state = self.optimizer.state.get(group['params'][0], None) if stored_state is not None: stored_state["exp_avg"] = torch.cat((stored_state["exp_avg"], torch.zeros_like(extension_tensor)), dim=0) stored_state["exp_avg_sq"] = torch.cat((stored_state["exp_avg_sq"], torch.zeros_like(extension_tensor)), dim=0) del self.optimizer.state[group['params'][0]] group["params"][0] = nn.Parameter(torch.cat((group["params"][0], extension_tensor), dim=0).requires_grad_(True)) self.optimizer.state[group['params'][0]] = stored_state optimizable_tensors[group["name"]] = group["params"][0] else: group["params"][0] = nn.Parameter(torch.cat((group["params"][0], extension_tensor), dim=0).requires_grad_(True)) optimizable_tensors[group["name"]] = group["params"][0] return optimizable_tensors def densification_postfix(self, new_xyz, new_knn_f, new_features_dc, new_features_rest, new_opacities, new_scaling, new_rotation, new_language_feature, new_instance_feature): d = {"xyz": new_xyz, "knn_f": new_knn_f, "f_dc": new_features_dc, "f_rest": new_features_rest, "opacity": new_opacities, "scaling" : new_scaling, "rotation" : new_rotation, "language_feature": new_language_feature, "instance_feature": new_instance_feature, } optimizable_tensors = self.cat_tensors_to_optimizer(d) self._xyz = optimizable_tensors["xyz"] self._knn_f = optimizable_tensors["knn_f"] self._features_dc = optimizable_tensors["f_dc"] self._features_rest = optimizable_tensors["f_rest"] self._opacity = optimizable_tensors["opacity"] self._scaling = optimizable_tensors["scaling"] self._rotation = optimizable_tensors["rotation"] self._language_feature = optimizable_tensors["language_feature"] self._instance_feature = optimizable_tensors["instance_feature"] self.xyz_gradient_accum = torch.zeros((self.get_xyz.shape[0], 1), device="cuda") self.xyz_gradient_accum_abs = torch.zeros((self.get_xyz.shape[0], 1), device="cuda") self.denom = torch.zeros((self.get_xyz.shape[0], 1), device="cuda") self.denom_abs = torch.zeros((self.get_xyz.shape[0], 1), device="cuda") self.max_radii2D = torch.zeros((self.get_xyz.shape[0]), device="cuda") self.max_weight = torch.zeros((self.get_xyz.shape[0]), device="cuda") def densify_and_split(self, grads, grad_threshold, grads_abs, grad_abs_threshold, scene_extent, max_radii2D, N=2): n_init_points = self.get_xyz.shape[0] # Extract points that satisfy the gradient condition padded_grad = torch.zeros((n_init_points), device="cuda") padded_grad[:grads.shape[0]] = grads.squeeze() padded_grads_abs = torch.zeros((n_init_points), device="cuda") padded_grads_abs[:grads_abs.shape[0]] = grads_abs.squeeze() padded_max_radii2D = torch.zeros((n_init_points), device="cuda") padded_max_radii2D[:max_radii2D.shape[0]] = max_radii2D.squeeze() selected_pts_mask = torch.where(padded_grad >= grad_threshold, True, False) selected_pts_mask = torch.logical_and(selected_pts_mask, torch.max(self.get_scaling, dim=1).values > self.percent_dense*scene_extent) if selected_pts_mask.sum() + n_init_points > self.max_all_points: limited_num = self.max_all_points - n_init_points padded_grad[~selected_pts_mask] = 0 ratio = limited_num / float(n_init_points) threshold = torch.quantile(padded_grad, (1.0-ratio)) selected_pts_mask = torch.where(padded_grad > threshold, True, False) # print(f"split {selected_pts_mask.sum()}, raddi2D {padded_max_radii2D.max()} ,{padded_max_radii2D.median()}") else: padded_grads_abs[selected_pts_mask] = 0 mask = (torch.max(self.get_scaling, dim=1).values > self.percent_dense*scene_extent) & (padded_max_radii2D > self.abs_split_radii2D_threshold) padded_grads_abs[~mask] = 0 selected_pts_mask_abs = torch.where(padded_grads_abs >= grad_abs_threshold, True, False) limited_num = min(self.max_all_points - n_init_points - selected_pts_mask.sum(), self.max_abs_split_points) if selected_pts_mask_abs.sum() > limited_num: ratio = limited_num / float(n_init_points) threshold = torch.quantile(padded_grads_abs, (1.0-ratio)) selected_pts_mask_abs = torch.where(padded_grads_abs > threshold, True, False) selected_pts_mask = torch.logical_or(selected_pts_mask, selected_pts_mask_abs) # print(f"split {selected_pts_mask.sum()}, abs {selected_pts_mask_abs.sum()}, raddi2D {padded_max_radii2D.max()} ,{padded_max_radii2D.median()}") stds = self.get_scaling[selected_pts_mask].repeat(N,1) means =torch.zeros((stds.size(0), 3),device="cuda") samples = torch.normal(mean=means, std=stds) rots = build_rotation(self._rotation[selected_pts_mask]).repeat(N,1,1) new_xyz = torch.bmm(rots, samples.unsqueeze(-1)).squeeze(-1) + self.get_xyz[selected_pts_mask].repeat(N, 1) new_scaling = self.scaling_inverse_activation(self.get_scaling[selected_pts_mask].repeat(N,1) / (0.8*N)) new_rotation = self._rotation[selected_pts_mask].repeat(N,1) new_features_dc = self._features_dc[selected_pts_mask].repeat(N,1,1) new_features_rest = self._features_rest[selected_pts_mask].repeat(N,1,1) new_opacity = self._opacity[selected_pts_mask].repeat(N,1) new_knn_f = self._knn_f[selected_pts_mask].repeat(N,1) new_language_feature = self._language_feature[selected_pts_mask].repeat(N,1) new_instance_feature = self._instance_feature[selected_pts_mask].repeat(N,1) self.densification_postfix(new_xyz, new_knn_f, new_features_dc, new_features_rest, new_opacity, new_scaling, new_rotation, new_language_feature, new_instance_feature) prune_filter = torch.cat((selected_pts_mask, torch.zeros(N * selected_pts_mask.sum(), device="cuda", dtype=bool))) self.prune_points(prune_filter) def densify_and_clone(self, grads, grad_threshold, scene_extent): n_init_points = self.get_xyz.shape[0] # Extract points that satisfy the gradient condition selected_pts_mask = torch.where(torch.norm(grads, dim=-1) >= grad_threshold, True, False) selected_pts_mask = torch.logical_and(selected_pts_mask, torch.max(self.get_scaling, dim=1).values <= self.percent_dense*scene_extent) if selected_pts_mask.sum() + n_init_points > self.max_all_points: limited_num = self.max_all_points - n_init_points grads_tmp = grads.squeeze().clone() grads_tmp[~selected_pts_mask] = 0 ratio = min(limited_num / float(n_init_points), 1) threshold = torch.quantile(grads_tmp, (1.0-ratio)) selected_pts_mask = torch.where(grads_tmp > threshold, True, False) if selected_pts_mask.sum() > 0: # print(f"clone {selected_pts_mask.sum()}") new_xyz = self._xyz[selected_pts_mask] stds = self.get_scaling[selected_pts_mask] means =torch.zeros((stds.size(0), 3),device="cuda") samples = torch.normal(mean=means, std=stds) rots = build_rotation(self._rotation[selected_pts_mask]) new_xyz = torch.bmm(rots, samples.unsqueeze(-1)).squeeze(-1) + self.get_xyz[selected_pts_mask] new_features_dc = self._features_dc[selected_pts_mask] new_features_rest = self._features_rest[selected_pts_mask] new_opacities = self._opacity[selected_pts_mask] new_scaling = self._scaling[selected_pts_mask] new_rotation = self._rotation[selected_pts_mask] new_knn_f = self._knn_f[selected_pts_mask] new_language_feature = self._language_feature[selected_pts_mask] new_instance_feature = self._instance_feature[selected_pts_mask] self.densification_postfix(new_xyz, new_knn_f, new_features_dc, new_features_rest, new_opacities, new_scaling, new_rotation, new_language_feature, new_instance_feature) def densify_and_prune(self, max_grad, abs_max_grad, min_opacity, extent, max_screen_size): grads = self.xyz_gradient_accum / self.denom grads_abs = self.xyz_gradient_accum_abs / self.denom_abs grads[grads.isnan()] = 0.0 grads_abs[grads_abs.isnan()] = 0.0 max_radii2D = self.max_radii2D.clone() self.densify_and_clone(grads, max_grad, extent) self.densify_and_split(grads, max_grad, grads_abs, abs_max_grad, extent, max_radii2D) prune_mask = (self.get_opacity < min_opacity).squeeze() if max_screen_size: big_points_vs = self.max_radii2D > max_screen_size big_points_ws = self.get_scaling.max(dim=1).values > 0.1 * extent prune_mask = torch.logical_or(torch.logical_or(prune_mask, big_points_vs), big_points_ws) self.prune_points(prune_mask) # print(f"all points {self._xyz.shape[0]}") torch.cuda.empty_cache() def add_densification_stats(self, viewspace_point_tensor, viewspace_point_tensor_abs, update_filter): self.xyz_gradient_accum[update_filter] += torch.norm(viewspace_point_tensor.grad[update_filter,:2], dim=-1, keepdim=True) self.xyz_gradient_accum_abs[update_filter] += torch.norm(viewspace_point_tensor_abs.grad[update_filter,:2], dim=-1, keepdim=True) self.denom[update_filter] += 1 self.denom_abs[update_filter] += 1 def get_points_depth_in_depth_map(self, fov_camera, depth, points_in_camera_space, scale=1): st = max(int(scale/2)-1,0) depth_view = depth[None,:,st::scale,st::scale] W, H = int(fov_camera.image_width/scale), int(fov_camera.image_height/scale) depth_view = depth_view[:H, :W] pts_projections = torch.stack( [points_in_camera_space[:,0] * fov_camera.Fx / points_in_camera_space[:,2] + fov_camera.Cx, points_in_camera_space[:,1] * fov_camera.Fy / points_in_camera_space[:,2] + fov_camera.Cy], -1).float()/scale mask = (pts_projections[:, 0] > 0) & (pts_projections[:, 0] < W) &\ (pts_projections[:, 1] > 0) & (pts_projections[:, 1] < H) & (points_in_camera_space[:,2] > 0.1) pts_projections[..., 0] /= ((W - 1) / 2) pts_projections[..., 1] /= ((H - 1) / 2) pts_projections -= 1 pts_projections = pts_projections.view(1, -1, 1, 2) map_z = torch.nn.functional.grid_sample(input=depth_view, grid=pts_projections, mode='bilinear', padding_mode='border', align_corners=True )[0, :, :, 0] return map_z, mask def get_points_from_depth(self, fov_camera, depth, scale=1): st = int(max(int(scale/2)-1,0)) depth_view = depth.squeeze()[st::scale,st::scale] rays_d = fov_camera.get_rays(scale=scale) depth_view = depth_view[:rays_d.shape[0], :rays_d.shape[1]] pts = (rays_d * depth_view[..., None]).reshape(-1,3) R = torch.tensor(fov_camera.R).float().cuda() T = torch.tensor(fov_camera.T).float().cuda() pts = (pts-T)@R.transpose(-1,-2) return pts def change_reqiures_grad(self, change, iteration, quiet=True): if change == "geometry": self._xyz.requires_grad_(True) self._knn_f.requires_grad_(True) self._features_dc.requires_grad_(True) self._features_rest.requires_grad_(True) self._scaling.requires_grad_(True) self._rotation.requires_grad_(True) self._opacity.requires_grad_(True) self.P.requires_grad_(True) self._language_feature.requires_grad_(False) self._instance_feature.requires_grad_(False) if not quiet: print(f'\n[ITER {iteration}] Training gaussian params') elif change == 'semantic': self._xyz.requires_grad_(True) self._knn_f.requires_grad_(True) self._features_dc.requires_grad_(True) self._features_rest.requires_grad_(True) self._scaling.requires_grad_(True) self._rotation.requires_grad_(True) self._opacity.requires_grad_(True) self.P.requires_grad_(True) self._language_feature.requires_grad_(True) self._instance_feature.requires_grad_(False) if not quiet: print(f'\n[ITER {iteration}] Training gaussian params and language feature') elif change == 'semantic_only': self._xyz.requires_grad_(False) self._knn_f.requires_grad_(False) self._features_dc.requires_grad_(False) self._features_rest.requires_grad_(False) self._scaling.requires_grad_(False) self._rotation.requires_grad_(False) self._opacity.requires_grad_(False) self.P.requires_grad_(False) self._language_feature.requires_grad_(True) self._instance_feature.requires_grad_(False) if not quiet: print(f'\n[ITER {iteration}] Training language feature') elif change == 'instance': self._xyz.requires_grad_(False) self._knn_f.requires_grad_(False) self._features_dc.requires_grad_(False) self._features_rest.requires_grad_(False) self._scaling.requires_grad_(False) self._rotation.requires_grad_(False) self._opacity.requires_grad_(False) self.P.requires_grad_(False) self._language_feature.requires_grad_(False) self._instance_feature.requires_grad_(True) if not quiet: print(f'\n[ITER {iteration}] Training instance feature') elif change == "pose_only": self._xyz.requires_grad_(False) self._knn_f.requires_grad_(False) self._features_dc.requires_grad_(False) self._features_rest.requires_grad_(False) self._scaling.requires_grad_(False) self._rotation.requires_grad_(False) self._opacity.requires_grad_(False) self.P.requires_grad_(True) self._language_feature.requires_grad_(False) self._instance_feature.requires_grad_(False) if not quiet: print(f'\n[ITER {iteration}] Training instance feature') elif change == 'finetune': self._xyz.requires_grad_(False) self._knn_f.requires_grad_(False) self._features_dc.requires_grad_(True) self._features_rest.requires_grad_(True) self._scaling.requires_grad_(False) self._rotation.requires_grad_(False) self._opacity.requires_grad_(False) self.P.requires_grad_(False) self._language_feature.requires_grad_(False) self._instance_feature.requires_grad_(False) if not quiet: print(f'\n[ITER {iteration}] finetune') else: raise ValueError('Unknown type!')