[Question]: v0.5格式的数据集下pointcloud.ply如何使用？

[XieWeikai]

Question

作者您好，我之前下载使用的是v0.1的数据集，其中每一条轨迹在一个独立的目录下，然后每一条轨迹都有对应的pointcloud.ply，里面根据颜色区分障碍物和path。最近我更新了数据集，使用v0.5的数据集，现在多条轨迹共用一个pointcloud.ply，请问是这多条轨迹的场景一样所以能共用吗？另外要如何区分不同的path呢？原先的navdp_dataset我看并没有更新，找不太到相关的代码😭。

[XieWeikai]

我打开点云看了下，点云里面是没有path了对吗？path就都由parquet里面取出的世界坐标来看了，和点云是同一个世界坐标系？

[wzcai99]

你好，目前path从点云中删除了，跟parquet里面的camera_trajectory是一个意思，另外关于dataloader的话基本只需要修改一下__init__部分，提供一段可以参考的代码:

def __init__(
        self,
        root_dirs,
        preload_path=False,
        memory_size=8,
        predict_size=24,
        batch_size=64,
        image_size=224,
        scene_data_scale=1.0,
        trajectory_data_scale=1.0,
        pixel_channel=7,
        action_dim=5,
        debug=False,
        preload=False,
        random_digit=False,
        prior_sample=False,
    ):

        self.dataset_dirs = np.array([p for p in os.listdir(root_dirs)])
        self.memory_size = memory_size
        self.image_size = image_size
        self.scene_scale_size = scene_data_scale
        self.trajectory_data_scale = trajectory_data_scale
        self.predict_size = predict_size
        self.action_dim = action_dim
        self.debug = debug

        self.trajectory_data_dir = []
        self.trajectory_rgb_path = []
        self.trajectory_depth_path = []
        self.trajectory_afford_path = []
        self.random_digit = random_digit
        self.prior_sample = prior_sample
        self.pixel_channel = pixel_channel
        self.item_cnt = 0
        self.batch_size = batch_size
        self.batch_time_sum = 0.0
        self._last_time = None

        if preload is False:
            for group_dir in self.dataset_dirs:  # gibson_zed, 3dfront ...                
                all_scene_dirs = np.array([p for p in os.listdir(os.path.join(root_dirs, group_dir))])
                select_scene_dirs = all_scene_dirs[
                    np.arange(0, all_scene_dirs.shape[0], 1 / self.scene_scale_size).astype(np.int32)
                ]
                
                for scene_dir in tqdm(select_scene_dirs):
                    chunk_name = os.listdir(os.path.join(root_dirs, group_dir, scene_dir, 'data'))[0]
                    data_dir = os.path.join(root_dirs, group_dir, scene_dir, f'data/{chunk_name}')
                    afford_dir = os.path.join(root_dirs, group_dir, scene_dir, 'meta/pointcloud.ply')
                    with jsonlines.open(os.path.join(root_dirs, group_dir, scene_dir, 'meta/episodes_stats.jsonl'), 'r') as reader:
                        episode_info = list(reader)
                    rgb_dir = os.path.join(root_dirs, group_dir, scene_dir, f"videos/{chunk_name}/observation.images.rgb/")
                    rgb_paths = [os.path.join(rgb_dir, p) for p in sorted(os.listdir(rgb_dir))]
                    
                    depth_dir = os.path.join(root_dirs, group_dir, scene_dir, f"videos/{chunk_name}/observation.images.depth/")
                    depth_paths = [os.path.join(depth_dir, p) for p in sorted(os.listdir(depth_dir))]
                    
                    data_paths = [os.path.join(data_dir, p) for p in sorted(os.listdir(data_dir))]
                    
                    for episode_idx, episode in enumerate(episode_info):
                        image_start_index = episode['image_index']['min']
                        image_end_index = episode['image_index']['max']
                        episode_rgb_path = np.array(rgb_paths)[image_start_index:image_end_index+1].tolist()
                        episode_depth_path = np.array(depth_paths)[image_start_index:image_end_index+1].tolist()
                        
                        try:
                            self.trajectory_data_dir.append(data_paths[episode_idx])
                            self.trajectory_rgb_path.append(episode_rgb_path)
                            self.trajectory_depth_path.append(episode_depth_path)
                            self.trajectory_afford_path.append(afford_dir)
                        except:
                            import pdb
                            pdb.set_trace()

            save_dict = {
                'trajectory_data_dir': self.trajectory_data_dir,
                'trajectory_rgb_path': self.trajectory_rgb_path,
                'trajectory_depth_path': self.trajectory_depth_path,
                'trajectory_afford_path': self.trajectory_afford_path,
            }
            with open(preload_path, 'w') as f:
                json.dump(save_dict, f, indent=4)
                
        else:
            load_dict = json.load(open(preload_path, 'r'))
            self.trajectory_data_dir = load_dict['trajectory_data_dir'] * 50
            self.trajectory_rgb_path = load_dict['trajectory_rgb_path'] * 50
            self.trajectory_depth_path = load_dict['trajectory_depth_path'] * 50
            self.trajectory_afford_path = load_dict['trajectory_afford_path'] * 50

[XieWeikai]

好的明白了！感谢！