import os import collections import json import random from copy import deepcopy import jsonlines from tqdm import tqdm import numpy as np import torch from .default import DATASET_REGISTRY from ..data_utils import build_rotate_mat from data.data_utils import (VICUNA_ACTION_TOKENS) from .scannet_base import ScanNetBase from scipy.spatial.transform import Rotation as R ONESTEPNAVI_ACTION_SPACE = { 'move_forward': 0, 'turn_left': 1, 'move_backward': 2, 'turn_right': 3, 'turn_left_forward': 4, 'turn_left_backward': 5, 'turn_right_backward': 6, 'turn_right_forward': 7, } ONESTEPNAVI_ACTION_SPACE_TOKENIZE = { k: v for k, v in zip(list(ONESTEPNAVI_ACTION_SPACE.values()), list(VICUNA_ACTION_TOKENS.keys())[:len(ONESTEPNAVI_ACTION_SPACE)]) } NAVI_ACTION_POOL = [ "What action should I take next step?", ] # one_data_for_training = { # "location" : location, # "orientation" : quaternion.tolist(), # "situation_multimodal" : situation_desp, # "situation_text" : situation_text, # "interaction" : interaction, # "instruction" : instruction, # "action" : { # "four_direction": [ # 1, # "turn left" # ], # "eight_direction": [ # 1, # "turn to foward left" # ], # "angle": 60.37389999185825 # }, # "meta_data" : one_data # } # mapping = { # 0 : "move forward", 1 : "turn left", 2 : "move backward", 3 : "turn right", 4 : "move forward" # } # mapping = { # 0 : "move forward", 1 : "turn to foward left", 2 : "turn left", 3 : "turn to back left", 4 : "move backward", # 5 : "turn to back right", 6 : "turn right", 7 : "turn to foward right", 8 : "move forward" # } @DATASET_REGISTRY.register() class ScanNetOneStepNavi(ScanNetBase): def __init__(self, cfg, split): super().__init__(cfg, split) self.dataset_cfg = cfg.data.next_step_navigation.args self.num_points = self.dataset_cfg.get('num_points', 1024) self.max_obj_len = self.dataset_cfg.get('max_obj_len', 60) self.pc_type = self.dataset_cfg.get('pc_type', 'gt') self.action_type = self.dataset_cfg.get('action_type', 'four_direction') self.modality_type = self.dataset_cfg.get('modality_type', 'multimodal') assert self.pc_type in ['gt', 'pred'] assert self.split in ['train', 'val', 'test'] if self.split == 'train': self.pc_type = 'gt' if self.split == 'test': self.split = 'val' self.action_mapping = { "four_direction": {0:0, 1:1, 2:2, 3:3, 4:0}, "eight_direction": {0:0, 2:1, 4:2, 6:3, 8:0, 1:4, 3:5, 5:6, 7:7}, } self.scan_ids = self._load_split(cfg, self.split) anno_file_path = os.path.join(cfg.data.msnn_base, 'msnn_scannet.json') with open(anno_file_path, 'r') as f: anno_info_all = json.load(f) print(f"Loading ScanNet ScanNetOneStepNavi {split}-set language") self.data, self.scan_ids = self._load_lang(anno_info_all, self.scan_ids) if cfg.debug.flag: self.data = self.data[:cfg.debug.debug_size] print(f"Finish loading ScanNetOneStepNavi {split}-set language") # load scans print(f"Loading ScanNet ScanNetOneStepNavi {split}-set scans") self.scan_data = self._load_scannet(self.scan_ids, self.pc_type, self.pc_type == 'gt') print(f"Finish loading ScanNet ScanNetOneStepNavi {split}-set data") def _load_lang(self, anno_info_all, select_scan_ids): output_list = [] scan_ids = [] for scan_id, samples_one_scene in anno_info_all.items(): if scan_id not in select_scan_ids: continue scan_ids.append(scan_id) for one_sample in samples_one_scene.values(): one_sample['insts'] = [int(x) for x in one_sample['insts']] output_list.append(one_sample) scan_ids = list(set(scan_ids)) return output_list, scan_ids def __len__(self): return len(self.data) # get inputs for scene encoder def preprocess_pcd(self, obj_pcds, return_anchor = False, rot_aug = True, situation = None): # rotate scene rot_matrix = build_rotate_mat(self.split, rot_aug=rot_aug) # normalize pc and calculate location obj_fts = [] obj_locs = [] for i, obj_pcd in enumerate(obj_pcds): if rot_matrix is not None: obj_pcd[:, :3] = np.matmul(obj_pcd[:, :3], rot_matrix.transpose()) obj_center = obj_pcd[:, :3].mean(0) obj_size = obj_pcd[:, :3].max(0) - obj_pcd[:, :3].min(0) obj_locs.append(np.concatenate([obj_center, obj_size], 0)) if return_anchor and i == 0: # Select a loc within the obj bbox as the anchor. anchor_loc = obj_pcd[:, :3].min(0) + np.random.rand(3) * obj_size # subsample pcd_idxs = np.random.choice(len(obj_pcd), size=self.num_points, replace=len(obj_pcd) < self.num_points) obj_pcd = obj_pcd[pcd_idxs] # normalize obj_pcd[:, :3] = obj_pcd[:, :3] - obj_pcd[:, :3].mean(0) max_dist = np.sqrt((obj_pcd[:, :3]**2).sum(1)).max() if max_dist < 1e-6: # take care of tiny point-clouds, i.e., padding max_dist = 1 obj_pcd[:, :3] = obj_pcd[:, :3] / max_dist obj_fts.append(obj_pcd) # convert to torch obj_fts = torch.from_numpy(np.stack(obj_fts, 0)) obj_locs = torch.from_numpy(np.array(obj_locs)) if return_anchor: anchor_loc = torch.from_numpy(anchor_loc) else: anchor_loc = torch.zeros(3).float() output_dict = { 'obj_fts': obj_fts, 'obj_locs': obj_locs, 'anchor_loc': anchor_loc, } if situation is not None: if rot_matrix is None: output_dict["situation"] = situation else: pos, ori = situation pos = np.array(pos) ori = np.array(ori) pos_new = pos.reshape(1, 3) @ rot_matrix.transpose() pos_new = pos_new.reshape(-1) ori_new = R.from_quat(ori).as_matrix() ori_new = rot_matrix @ ori_new ori_new = R.from_matrix(ori_new).as_quat() ori_new = ori_new.reshape(-1) output_dict["situation"] = (pos_new, ori_new) return output_dict # get inputs for scene encoder def _get_scene_encoder_input(self, obj_pcds, scan_insts, situation = None): # Dict: { int: np.ndarray (N, 6) } if len(obj_pcds) <= self.max_obj_len: # Dict to List selected_obj_pcds = list(obj_pcds.values()) else: # crop objects to max_obj_len selected_obj_pcds = [] # select relevant objs first for i in scan_insts: if i in obj_pcds: selected_obj_pcds.append(obj_pcds[i]) num_selected_objs = len(selected_obj_pcds) if num_selected_objs >= self.max_obj_len: random.shuffle(selected_obj_pcds) selected_obj_pcds = selected_obj_pcds[:self.max_obj_len] else: # select from remaining objs remained_obj_idx = [i for i in obj_pcds.keys() if i not in scan_insts] random.shuffle(remained_obj_idx) for i in remained_obj_idx[: self.max_obj_len - num_selected_objs]: selected_obj_pcds.append(obj_pcds[i]) assert len(selected_obj_pcds) == self.max_obj_len output_dict = self.preprocess_pcd(selected_obj_pcds, return_anchor = False, rot_aug = True, situation = situation) return output_dict def __getitem__(self, index): one_sample = self.data[index] if self.modality_type == 'multimodal': situation = one_sample['situation_multimodal'] else: situation = one_sample['situation_text'] interaction = one_sample['interaction'] anchor_loc = one_sample['location'] anchor_orientation = one_sample['orientation'] question = random.choice(NAVI_ACTION_POOL) question = interaction + " " + question # load scene data scan_id = one_sample['scan_id'] obj_pcds = self.scan_data[scan_id]['obj_pcds'] obj_pcds = {int(k): obj_pcds[k] for k in range(len(obj_pcds))} action_token_list = [] action_gt_code = one_sample['action'][self.action_type][0] action_gt_code = self.action_mapping[self.action_type][action_gt_code] action_gt = ONESTEPNAVI_ACTION_SPACE_TOKENIZE[action_gt_code] action_token_list.append(action_gt) action_text_list = [] action_text = one_sample['action'][self.action_type][1] action_text_list.append(action_text) ### scene input #### output_dict = self._get_scene_encoder_input(obj_pcds, one_sample['insts'], situation = (anchor_loc, anchor_orientation)) obj_fts = output_dict['obj_fts'] obj_locs = output_dict['obj_locs'] anchor_loc, anchor_orientation = output_dict["situation"] data_dict = { "situation": situation, "situation_pos": np.array(anchor_loc), "situation_rot": np.array(anchor_orientation), "question": question, "action_token_list": action_token_list, "action_text_list": action_text_list, "obj_fts": obj_fts, "obj_locs": obj_locs, "scan_id": scan_id, } return data_dict