# Copyright (c) Meta Platforms, Inc. and affiliates. # All rights reserved. # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. # -------------------------------------------------------- # References: # NoMaD, GNM, ViNT: https://github.com/robodhruv/visualnav-transformer # -------------------------------------------------------- # 这个版本为加入了相机编码版本,并且把min_goal_dist,max_goal_dist分别约束到-16, 16 import cv2 import numpy as np import torch import os from PIL import Image from typing import Tuple import yaml import pickle import tqdm from torch.utils.data import Dataset from misc import angle_difference, get_data_path, get_delta_np, normalize_data, to_local_coords from project_functions import AirsimCoordsConverter, reproject_depth_to_other_pose_2seq, project_to_2d_image_2seq, resize_image_half class BaseDataset(Dataset): def __init__( self, data_folder: str, data_split_folder: str, dataset_name: str, image_size: Tuple[int, int], min_dist_cat: int, max_dist_cat: int, len_traj_pred: int, traj_stride: int, context_size: int, transform: object, traj_names: str, normalize: bool = True, predefined_index: list = None, goals_per_obs: int = 1, ): self.data_folder = data_folder self.data_split_folder = data_split_folder self.dataset_name = dataset_name self.goals_per_obs = goals_per_obs if 'airvln' in dataset_name: self.coords_converter = AirsimCoordsConverter() traj_names_file = os.path.join(data_split_folder, traj_names) with open(traj_names_file, "r") as f: file_lines = f.read() self.traj_names = file_lines.split("\n") if "" in self.traj_names: self.traj_names.remove("") self.image_size = image_size self.distance_categories = list(range(min_dist_cat, max_dist_cat + 1)) self.min_dist_cat = self.distance_categories[0] self.max_dist_cat = self.distance_categories[-1] self.len_traj_pred = len_traj_pred self.traj_stride = traj_stride self.context_size = context_size self.normalize = normalize # load data/data_config.yaml with open("config/data_config.yaml", "r") as f: all_data_config = yaml.safe_load(f) dataset_names = list(all_data_config.keys()) dataset_names.sort() # use this index to retrieve the dataset name from the data_config.yaml self.data_config = all_data_config[self.dataset_name] self.transform = transform self._load_index(predefined_index) self.ACTION_STATS = {} for key in all_data_config['action_stats']: self.ACTION_STATS[key] = np.expand_dims(all_data_config['action_stats'][key], axis=0) def _load_index(self, predefined_index) -> None: """ Generates a list of tuples of (obs_traj_name, goal_traj_name, obs_time, goal_time) for each observation in the dataset """ if predefined_index: print(f"****** Using a predefined evaluation index... {predefined_index}******") with open(predefined_index, "rb") as f: self.index_to_data = pickle.load(f) return else: print("****** Evaluating from NON PREDEFINED index... ******") index_to_data_path = os.path.join( self.data_split_folder, f"dataset_dist_{self.min_dist_cat}_to_{self.max_dist_cat}_n{self.context_size}_len_traj_pred_{self.len_traj_pred}.pkl", ) self.index_to_data, self.goals_index = self._build_index() with open(index_to_data_path, "wb") as f: pickle.dump((self.index_to_data, self.goals_index), f) def _build_index(self, use_tqdm: bool = False): """ Build an index consisting of tuples (trajectory name, time, max goal distance) """ samples_index = [] goals_index = [] for traj_name in tqdm.tqdm(self.traj_names, disable=not use_tqdm, dynamic_ncols=True): traj_data = self._get_trajectory(traj_name) traj_len = len(traj_data["position"]) for goal_time in range(0, traj_len): goals_index.append((traj_name, goal_time)) begin_time = self.context_size - 1 end_time = traj_len - self.len_traj_pred for curr_time in range(begin_time, end_time, self.traj_stride): max_goal_distance = min(self.max_dist_cat, traj_len - curr_time - 1) min_goal_distance = max(self.min_dist_cat, -curr_time) samples_index.append((traj_name, curr_time, min_goal_distance, max_goal_distance)) return samples_index, goals_index def _get_trajectory(self, trajectory_name): with open(os.path.join(self.data_folder, trajectory_name, "traj_data.pkl"), "rb") as f: traj_data = pickle.load(f) for k,v in traj_data.items(): traj_data[k] = v.astype('float') return traj_data def __len__(self) -> int: return len(self.index_to_data) def _compute_projected_image(self, traj_data, curr_time, goal_time, rgb_img): pose_src = traj_data["pose"][curr_time] pose_dst = traj_data["pose"][goal_time] depth_map = traj_data["depth"][curr_time] K = traj_data["K"] # print(f"pose_src shape: {pose_src.shape}, pose_dst shape: {pose_dst.shape}, dep_map shape: {depth_map.shape}") projected_images = self.generate_augmented_image(K=K, depth_map=depth_map, rgb_img=rgb_img, pose_src=pose_src, pose_dst=pose_dst) # projected_images = self.generate_augmented_image_v2(K=K, depth_map=depth_map, rgb_img=rgb_img, pose_src=pose_src, pose_dst=pose_dst) return projected_images def generate_augmented_image(self, K, depth_map, rgb_img, pose_src, pose_dst) -> np.ndarray: """ 基于深度图 + pose 生成从另一个相机视角观察到的图像。 """ image_size = depth_map.shape # (H, W) if rgb_img.shape[:2] != image_size: rgb_img = resize_image_half(rgb_img) points_3d, colors = reproject_depth_to_other_pose_2seq(K, depth_map, rgb_img, pose_src, pose_dst) images = project_to_2d_image_2seq(K, points_3d, colors, image_size) # (H, W, 3, goal_time) return images def generate_augmented_image_v2(self, K, depth_map, rgb_img, pose_src, pose_dst): goal_times = pose_dst.shape[0] images = [] for i in range(goal_times): p_dst = pose_dst[i] images.append(self.warpPerspective(K, depth_map, rgb_img, pose_src, p_dst)) return images def warpPerspective(self, K, depth_map, rgb_img, pose_src, pose_dst, out_size=None, fill_value=0) -> np.ndarray: """ 重投影 src 图像到 dst 相机视角（同内参 K）。 参数 ---- K : (3,3) numpy.ndarray 相机内参（src/dst 相同） depth_map : (H,W) numpy.ndarray 与 rgb_img 对齐的深度（单位米） rgb_img : (H,W,3) uint8 源图像（相机 src 拍摄） pose_src : (4,4) numpy.ndarray 相机 src 的 camera-to-world (c2w) 位姿 pose_dst : (4,4) numpy.ndarray 相机 dst 的 camera-to-world (c2w) 位姿 out_size : (W_out, H_out) or None 目标图像尺寸；None 时用源图像尺寸 fill_value : int or tuple 空洞填充值（背景） 返回 ---- img_dst : (H_out, W_out, 3) uint8 在 dst 视角下渲染的图像 z_dst : (H_out, W_out) float32 目标视角的深度（可用于可视化/调试） """ H, W = depth_map.shape if out_size is None: W_out, H_out = W, H else: W_out, H_out = out_size # 1) 构造像素网格（源图像） u, v = np.meshgrid(np.arange(W), np.arange(H)) # (H,W) ones = np.ones_like(u, dtype=np.float32) pix_src_h = np.stack([u, v, ones], axis=-1).reshape(-1, 3).T # 3xN depth = depth_map.reshape(-1).astype(np.float32) # N # 2) 反投影到 cam_src 坐标： x_src = depth * K^{-1} * [u,v,1]^T Kinv = np.linalg.inv(K) x_src = (Kinv @ pix_src_h) * depth # 3xN # 3) cam_src -> world -> cam_dst # T_src2dst = (pose_dst)^{-1} @ pose_src T_src2dst = np.linalg.inv(pose_dst) @ pose_src R = T_src2dst[:3, :3] t = T_src2dst[:3, 3:4] # 3x1 x_dst = (R @ x_src) + t # 3xN # 4) 投影到 dst 像素 pix_dst_h = K @ x_dst # 3xN z = pix_dst_h[2, :] + 1e-6 u_dst = (pix_dst_h[0, :] / z) v_dst = (pix_dst_h[1, :] / z) # 5) 前向“光栅化” + Z-buffer（处理遮挡；最近深度覆盖） img_dst = np.full((H_out, W_out, 3), fill_value, dtype=rgb_img.dtype) z_dst = np.full((H_out, W_out), np.inf, dtype=np.float32) # 只保留落在目标画幅内的点 u_round = np.round(u_dst).astype(np.int64) v_round = np.round(v_dst).astype(np.int64) valid = ( (z > 0) & (u_round >= 0) & (u_round < W_out) & (v_round >= 0) & (v_round < H_out) & (depth > 0) ) src_colors = rgb_img.reshape(-1, 3)[valid] uu = u_round[valid] vv = v_round[valid] zz = z[valid].astype(np.float32) # Z-buffer：对同一像素，保留深度更小（更近）的样本 # 用扁平索引实现原子“最小深度写入” flat_idx = vv * W_out + uu # 为每个像素找到最小深度的索引 order = np.argsort(zz) # 从近到远 flat_idx = flat_idx[order] zz = zz[order] src_colors = src_colors[order] # 只保留每个像素第一次出现（即最小深度） _, first_pos = np.unique(flat_idx, return_index=True) keep = np.zeros_like(order, dtype=bool) keep[first_pos] = True flat_idx = flat_idx[keep] zz = zz[keep] src_colors = src_colors[keep] # 写入帧缓冲 z_dst.flat[flat_idx] = zz img_dst.reshape(-1, 3)[flat_idx] = src_colors # 补全` # hole_mask = (img_dst.mean(axis=2) == 0).astype(np.uint8) # print(hole_mask.sum()) # hole_mask = cv2.dilate(hole_mask, np.ones((3,3), np.uint8), iterations=1) # img_dst = cv2.inpaint(img_dst, hole_mask*255, inpaintRadius=3, flags=cv2.INPAINT_TELEA) return img_dst def _compute_actions(self, traj_data, curr_time, goal_time, rgb_img): start_index = curr_time end_index = curr_time + self.len_traj_pred + 1 yaw = traj_data["yaw"][start_index:end_index] positions = traj_data["point"][start_index:end_index] goal_pos = traj_data["point"][goal_time] goal_yaw = traj_data["yaw"][goal_time] if len(yaw.shape) == 2: yaw = yaw.squeeze(1) if yaw.shape != (self.len_traj_pred + 1,): raise ValueError("is used?") # const_len = self.len_traj_pred + 1 - yaw.shape[0] # yaw = np.concatenate([yaw, np.repeat(yaw[-1], const_len)]) # positions = np.concatenate([positions, np.repeat(positions[-1][None], const_len, axis=0)], axis=0) waypoints_pos = to_local_coords(positions, positions[0], yaw[0]) waypoints_yaw = angle_difference(yaw[0], yaw) actions = np.concatenate([waypoints_pos, waypoints_yaw.reshape(-1, 1)], axis=-1) actions = actions[1:] goal_pos = to_local_coords(goal_pos, positions[0], yaw[0]) goal_yaw = angle_difference(yaw[0], goal_yaw) if self.normalize: actions[:, :3] /= self.data_config["metric_waypoint_spacing"] goal_pos[:, :3] /= self.data_config["metric_waypoint_spacing"] goal_pos = np.concatenate([goal_pos, goal_yaw.reshape(-1, 1)], axis=-1) projected_images = self._compute_projected_image(traj_data, curr_time, goal_time, rgb_img) return actions, goal_pos, projected_images class TrainingDataset(BaseDataset): def __init__( self, data_folder: str, data_split_folder: str, dataset_name: str, image_size: Tuple[int, int], min_dist_cat: int, max_dist_cat: int, len_traj_pred: int, traj_stride: int, context_size: int, transform: object, traj_names: str = 'traj_names.txt', normalize: bool = True, predefined_index: list = None, goals_per_obs: int = 1, ): super().__init__(data_folder, data_split_folder, dataset_name, image_size, min_dist_cat, max_dist_cat, len_traj_pred, traj_stride, context_size, transform, traj_names, normalize, predefined_index, goals_per_obs) def __getitem__(self, i: int) -> Tuple[torch.Tensor]: try: f_curr, curr_time, min_goal_dist, max_goal_dist = self.index_to_data[i] goal_offset = np.random.randint(min_goal_dist//4, max_goal_dist//4 + 1, size=(self.goals_per_obs)) goal_time = (curr_time + goal_offset).astype('int') rel_time = (goal_offset).astype('float')/(128.) # TODO: refactor, currently a fixed const context_times = list(range(curr_time - self.context_size + 1, curr_time + 1)) context = [(f_curr, t) for t in context_times] + [(f_curr, t) for t in goal_time] context_t = [t for _, t in context] # print(f"curr_time: {curr_time}, context_size: {self.context_size}, min_goal_dist: {min_goal_dist}, max_goal_dist: {max_goal_dist}") obs_image = torch.stack([self.transform(Image.open(get_data_path(self.data_folder, f, t))) for f, t in context]) # Load other trajectory data curr_traj_data = self._get_trajectory(f_curr) # print(f"traj K: {curr_traj_data['K']}, position: {curr_traj_data['position']}, pose: {curr_traj_data['pose']}, point: {curr_traj_data['point']}") # aug f_img, t_img = context[self.context_size-1] # curr_time img # print(f"context: {context}, f_img: {f_img}, t_img: {t_img}, curr time: {curr_time}") rgb_img = cv2.imread(get_data_path(self.data_folder, f_img, t_img)) rgb_img = cv2.cvtColor(rgb_img, cv2.COLOR_BGR2RGB) # Compute actions _, goal_pos, projected_images = self._compute_actions(curr_traj_data, curr_time, goal_time, rgb_img) goal_pos[:, :3] = normalize_data(goal_pos[:, :3], self.ACTION_STATS) projected_tensor_list = [self.transform(Image.fromarray(img)) for img in projected_images] projected_tensor = torch.stack(projected_tensor_list, dim=0) # Compute camera mats T_wc = curr_traj_data['pose'][context_t] T_wc = torch.as_tensor(T_wc, dtype=torch.float32) T_cw = torch.linalg.inv(T_wc) # print(f"t cw shape: {T_cw.size()}, context_t: {context_t}") # points = [curr_traj_data['point'][t] for _, t in context] # oriens = [[curr_traj_data['pitch'][t], curr_traj_data['roll'][t], curr_traj_data['yaw'][t]] for _, t in context] # poses = np.concatenate((np.array(points), np.array(oriens)), axis=-1) # t_cws = torch.as_tensor( # np.array([self.coords_converter.trans_cam2world(poses[tt]) for tt in range(len(poses))]), dtype=torch.float32 # ) # print("t_cws: ", t_cws - T_cw) # ori_poses = np.array([curr_traj_data['pose'][t] for _, t in context]) # print("ori_pose: ", ori_poses) # # ===================== 保存图像 ===================== # vis_root = './visualizations' # sample_dir = os.path.join(vis_root, f'airvln_16_v2', f'sample_{i}') # os.makedirs(sample_dir, exist_ok=True) # # 1. 保存 curr_frame # curr_img_save_path = os.path.join(sample_dir, f'curr_frame_{curr_time}.png') # Image.fromarray(rgb_img).save(curr_img_save_path) # # 2. 保存 goal_frame # for idx, (f_curr, t_goal) in enumerate(context[self.context_size:]): # goal_img_path = get_data_path(self.data_folder, f_curr, t_goal) # goal_img = Image.open(goal_img_path) # goal_img.save(os.path.join(sample_dir, f'goal_{idx}_{t_goal}.png')) # # 3. 保存 projected goal frame # for idx, proj_img in enumerate(projected_images): # proj_img_save_path = os.path.join(sample_dir, f'projected_goal_{idx}.png') # Image.fromarray(proj_img).save(proj_img_save_path) # # ==================================================== return ( torch.as_tensor(obs_image, dtype=torch.float32), torch.as_tensor(goal_pos, dtype=torch.float32), torch.as_tensor(rel_time, dtype=torch.float32), torch.as_tensor(projected_tensor, dtype=torch.float32), torch.as_tensor(T_cw, dtype=torch.float32), ) except Exception as e: print(f"Exception in {self.dataset_name}", e) raise Exception(e) class EvalDataset(BaseDataset): def __init__( self, data_folder: str, data_split_folder: str, dataset_name: str, image_size: Tuple[int, int], min_dist_cat: int, max_dist_cat: int, len_traj_pred: int, traj_stride: int, context_size: int, transform: object, traj_names: str, normalize: bool = True, predefined_index: list = None, goals_per_obs: int = 1, ): super().__init__(data_folder, data_split_folder, dataset_name, image_size, min_dist_cat, max_dist_cat, len_traj_pred, traj_stride, context_size, transform, traj_names, normalize, predefined_index, goals_per_obs) def __getitem__(self, i: int) -> Tuple[torch.Tensor]: try: f_curr, curr_time, _, _ = self.index_to_data[i] context_times = list(range(curr_time - self.context_size + 1, curr_time + 1)) pred_times = list(range(curr_time + 1, curr_time + self.len_traj_pred + 1)) context = [(f_curr, t) for t in context_times] pred = [(f_curr, t) for t in pred_times] obs_image = torch.stack([self.transform(Image.open(get_data_path(self.data_folder, f, t))) for f, t in context]) pred_image = torch.stack([self.transform(Image.open(get_data_path(self.data_folder, f, t))) for f, t in pred]) curr_traj_data = self._get_trajectory(f_curr) # Compute last rgb image f_img, t_img = context[-1] # curr_time img rgb_img = cv2.imread(get_data_path(self.data_folder, f_img, t_img)) rgb_img = cv2.cvtColor(rgb_img, cv2.COLOR_BGR2RGB) # Compute actions actions, _, projected_images = self._compute_actions(curr_traj_data, curr_time, np.array(pred_times), rgb_img) # last argument is dummy goal actions[:, :3] = normalize_data(actions[:, :3], self.ACTION_STATS) delta = get_delta_np(actions) # Compute projected tensor projected_tensor_list = [self.transform(Image.fromarray(img)) for img in projected_images] projected_tensor = torch.stack(projected_tensor_list, dim=0) print(f"Step {i} has been projected.") print(f"projected_images shape: {projected_images.shape}, projected_tensor shape: {projected_tensor.size()}") # Compute T T_wc_ctx = curr_traj_data['pose'][context_times] # (context_size, 4, 4) T_wc_pred = curr_traj_data['pose'][pred_times] # (len_traj_pred, 4, 4) T_cw_ctx = torch.linalg.inv(torch.as_tensor(T_wc_ctx, dtype=torch.float32)) # (context_size, 4, 4) T_cw_pred = torch.linalg.inv(torch.as_tensor(T_wc_pred, dtype=torch.float32)) # (len_traj_pred, 4, 4) return ( torch.tensor([i], dtype=torch.float32), # for logging purposes torch.as_tensor(obs_image, dtype=torch.float32), torch.as_tensor(pred_image, dtype=torch.float32), torch.as_tensor(delta, dtype=torch.float32), torch.as_tensor(projected_tensor, dtype=torch.float32), torch.as_tensor(T_cw_ctx, dtype=torch.float32), #obs torch.as_tensor(T_cw_pred, dtype=torch.float32), #pred ) except Exception as e: print(f"Exception in {self.dataset_name}", e) raise Exception(e) class TrajectoryEvalDataset(BaseDataset): def __init__( self, data_folder: str, data_split_folder: str, dataset_name: str, image_size: Tuple[int, int], min_dist_cat: int, max_dist_cat: int, len_traj_pred: int, traj_stride: int, context_size: int, transform: object, traj_names: str, normalize: bool = True, predefined_index: list = None, goals_per_obs: int = 1, ): super().__init__(data_folder, data_split_folder, dataset_name, image_size, min_dist_cat, max_dist_cat, len_traj_pred, traj_stride, context_size, transform, traj_names, normalize, predefined_index, goals_per_obs) def _sample_goal(self, trajectory_name, curr_time, min_goal_dist, max_goal_dist): """ Sample a goal from the future in the same trajectory. Returns: (trajectory_name, goal_time, goal_is_negative) """ goal_offset = np.random.randint(min_goal_dist, max_goal_dist + 1) goal_time = curr_time + int(goal_offset) return trajectory_name, goal_time, False def __getitem__(self, i: int) -> Tuple[torch.Tensor]: try: f_curr, curr_time, min_goal_dist, max_goal_dist = self.index_to_data[i] f_goal, goal_time, _ = self._sample_goal(f_curr, curr_time, min_goal_dist, max_goal_dist) context_times = list(range(curr_time - self.context_size + 1, curr_time + 1)) context = [(f_curr, t) for t in context_times] obs_image = torch.stack([self.transform(Image.open(get_data_path(self.data_folder, f, t))) for f, t in context]) goal_image = self.transform(Image.open(get_data_path(self.data_folder, f_goal, goal_time))).unsqueeze(0) curr_traj_data = self._get_trajectory(f_curr) # Compute actions, goal_pos, projected images f_img, t_img = context[-1] # curr_time img rgb_img = cv2.imread(get_data_path(self.data_folder, f_img, t_img)) rgb_img = cv2.cvtColor(rgb_img, cv2.COLOR_BGR2RGB) actions, goal_pos, projected_images = self._compute_actions(curr_traj_data, curr_time, np.array([goal_time]), rgb_img) projected_tensor_list = [self.transform(Image.fromarray(img)) for img in projected_images] projected_tensor = torch.stack(projected_tensor_list, dim=0) T_wc_ctx = curr_traj_data['pose'][context_times] T_cw_ctx = torch.linalg.inv(torch.as_tensor(T_wc_ctx, dtype=torch.float32)) T_wc_goal = curr_traj_data['pose'][[goal_time]] T_cw_goal = torch.linalg.inv(torch.as_tensor(T_wc_goal, dtype=torch.float32)) return ( torch.tensor([i], dtype=torch.float32), # for logging purposes torch.as_tensor(obs_image, dtype=torch.float32), torch.as_tensor(goal_image, dtype=torch.float32), torch.as_tensor(actions, dtype=torch.float32), torch.as_tensor(goal_pos, dtype=torch.float32), torch.as_tensor(projected_tensor, dtype=torch.float32), torch.as_tensor(T_cw_ctx, dtype=torch.float32), torch.as_tensor(T_cw_goal, dtype=torch.float32), ) except Exception as e: print(f"Exception in {self.dataset_name}", e) raise Exception(e)