# Copyright 2022 The Kubric Authors # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # https://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Worker file for the Multi-Object Video (MOVi) C (and CC) datasets. * The number of objects is randomly chosen between --min_num_objects (3) and --max_num_objects (10) * The objects are randomly chosen from the Google Scanned Objects dataset * Background is an random HDRI from the HDRI Haven dataset, projected onto a Dome (half-sphere). The HDRI is also used for lighting the scene. """ import logging import bpy import copy import os import kubric as kb from kubric.simulator import PyBullet from kubric.renderer import Blender import numpy as np import random import shutil from GSO_transfer import GSO_dict from utils import save_scene_instruction, dataset_dir # --- Some configuration values DATASET_TYPE = "closer" # the region in which to place objects [(min), (max)] SPAWN_REGION = [(-8, -8, 0), (8, 8, 5)] SPAWN_REGION_OBJ = [[-6, -6, 0.5], [6, 6, 0.5]] VELOCITY_RANGE = [(-4., -4., 0.), (4., 4., 0.)] # --- CLI arguments parser = kb.ArgumentParser() parser.add_argument("--objects_split", choices=["train", "test"], default="train") # Configuration for the objects of the scene parser.add_argument("--min_num_objects", type=int, default=1, help="minimum number of objects") parser.add_argument("--max_num_objects", type=int, default=5, help="maximum number of objects") # Configuration for the floor and background parser.add_argument("--floor_friction", type=float, default=0.3) parser.add_argument("--floor_restitution", type=float, default=0.5) parser.add_argument("--backgrounds_split", choices=["train", "test"], default="train") parser.add_argument("--camera", choices=["fixed_random", "linear_movement"], default="fixed_random") parser.add_argument("--max_camera_movement", type=float, default=4.0) parser.add_argument("--smallest_scale", type=float, default=2.) parser.add_argument("--largest_scale", type=float, default=4.) # Configuration for the source of the assets parser.add_argument("--kubasic_assets", type=str, default="gs://kubric-public/assets/KuBasic/KuBasic.json") parser.add_argument("--hdri_assets", type=str, default="gs://kubric-public/assets/HDRI_haven/HDRI_haven.json") parser.add_argument("--gso_assets", type=str, default="gs://kubric-public/assets/GSO/GSO.json") parser.add_argument("--save_state", dest="save_state", action="store_true") parser.set_defaults(save_state=False, frame_end=24, frame_rate=12, resolution=512) parser.add_argument("--sub_outputdir", type=str, default="test sub output dir") parser.add_argument("--generate_idx", type=int, default=-1, help="generation idx") FLAGS = parser.parse_args() import pyquaternion as pyquat def default_rng(): return np.random.RandomState() def random_rotation(axis=None, rng=default_rng()): """ Compute a random rotation as a quaternion. If axis is None the rotation is sampled uniformly over all possible orientations. Otherwise it corresponds to a random rotation around the given axis.""" if axis is None: # uniform across rotation space # copied from pyquat.Quaternion.random to be able to use a custom rng r1, r2, r3 = rng.random(3) q1 = np.sqrt(1.0 - r1) * (np.sin(2 * np.pi * r2)) q2 = np.sqrt(1.0 - r1) * (np.cos(2 * np.pi * r2)) q3 = np.sqrt(r1) * (np.sin(2 * np.pi * r3)) q4 = np.sqrt(r1) * (np.cos(2 * np.pi * r3)) return q1, q2, q3, q4 else: if isinstance(axis, str) and axis.upper() in ["X", "Y", "Z"]: axis = {"X": (1., 0., 0.), "Y": (0., 1., 0.), "Z": (0., 0., 1.)}[axis.upper()] # quat = pyquat.Quaternion(axis=axis, angle=rng.uniform(0, 2*np.pi)) quat = pyquat.Quaternion(axis=axis, angle=rng.uniform(-0.5*np.pi, 0.5*np.pi)) # -0.5pi -- 0.5pi return tuple(quat) from kubric.core import objects def rotation_sampler(axis=None): def _sampler(obj: objects.PhysicalObject, rng): obj.quaternion = random_rotation(axis=axis, rng=rng) return _sampler def move_until_no_overlap(asset, simulator, spawn_region=((-1, -1, -1), (1, 1, 1)), max_trials=100, rng=default_rng()): return kb.randomness.resample_while(asset, # samplers=[rotation_sampler(axis='Z'), kb.randomness.position_sampler(spawn_region)], samplers=[kb.randomness.position_sampler(spawn_region)], condition=simulator.check_overlap, max_trials=max_trials, rng=rng) def check_ok(obj, pos, region): # import pdb; pdb.set_trace() x, y, z = pos if pos[0]region[1][0] or pos[1]region[1][1]: #or pos[2]region[1][2]: return False if simulator.check_overlap(obj): return False return True def get_obj_x_left(bound, scale): return -bound[0][0] * scale[0] def get_obj_x_right(bound, scale): return bound[1][0] * scale[0] def get_obj_y_front(bound, scale): return -bound[0][1] * scale[1] def get_obj_y_behind(bound, scale): return bound[1][1] * scale[1] def get_obj_z(bound, scale): return bound[0][2] * scale[2] def get_obj_z_up(bound, scale): return bound[1][2] * scale[2] # def get_new_pos(bounds, scale, ref_location, ref_pos, ref_z_up, ref_object, rng): # obj_z = - get_obj_z(bounds, scale) # # import pdb; pdb.set_trace() # ref_x_left, ref_x_right, ref_y_front, ref_y_behind = get_obj_x_left(ref_object.bounds, ref_object.scale), get_obj_x_right(ref_object.bounds, ref_object.scale), get_obj_y_front(ref_object.bounds, ref_object.scale), get_obj_y_behind(ref_object.bounds, ref_object.scale) # ref_x, ref_y, ref_z = ref_pos # if ref_location == 'front': # return [rng.uniform(ref_x-0.5, ref_x+0.5), rng.uniform(ref_y-ref_y_front-6, ref_y-ref_y_front-2), obj_z+0.02] # elif ref_location == 'behind': # return [rng.uniform(ref_x-0.5, ref_x+0.5), rng.uniform(ref_y+ref_y_behind+2, ref_y+ref_y_behind+6), obj_z+0.02] # elif ref_location == 'left': # return [rng.uniform(ref_x-ref_x_left-6, ref_x-ref_x_left-2), rng.uniform(ref_y-0.5, ref_y+0.5), obj_z+0.02] # elif ref_location == 'right': # return [rng.uniform(ref_x+ref_x_right+2, ref_x+ref_x_right+6), rng.uniform(ref_y-0.5, ref_y+0.5), obj_z+0.02] # elif ref_location == 'on': # return [ref_x, ref_y, ref_z+ref_z_up+obj_z+1] def get_new_pos(bounds, scale, ref_location, ref_pos, ref_z_up, ref_object, rng): # Calculate the z-position based on the object's bounds and scale obj_z = -get_obj_z(bounds, scale) + 0.2 # Ensuring the object is slightly above the ground # Extract the bounds from the SPAWN_REGION_OBJ variable spawn_x_min, spawn_y_min, _ = SPAWN_REGION_OBJ[0] spawn_x_max, spawn_y_max, _ = SPAWN_REGION_OBJ[1] # Calculate the maximum offsets within the given spawn bounds max_offset_x = spawn_x_max - spawn_x_min + 1 max_offset_y = spawn_y_max - spawn_y_min + 1 # Define absolute positions for each location using straightforward calculations import random # Return the position for the specified location return locations.get(ref_location, [ref_pos[0], ref_pos[1], obj_z]) # Default position if location is not specified def add_new_obj(scene, new_obj, ref_location, ref_object, rng, max_trails=50): ref_obj_pos = ref_object.position # import pdb; pdb.set_trace() ref_obj_z_up = get_obj_z_up(ref_object.bounds, ref_object.scale) new_obj_pos = get_new_pos(new_obj.bounds, new_obj.scale, ref_location, ref_obj_pos, ref_obj_z_up, ref_object, rng) new_obj.position = new_obj_pos scene += new_obj # import pdb; pdb.set_trace() trails = 0 while not check_ok(new_obj, new_obj.position, SPAWN_REGION_OBJ): trails += 1 # import pdb; pdb.set_trace() new_obj.position = get_new_pos(new_obj.bounds, new_obj.scale, ref_location, ref_obj_pos, ref_obj_z_up, ref_object, rng) # new_obj.quaternion = random_rotation(axis="Z", rng=rng) if trails > max_trails: print('cannot put the object, break') # import pdb; pdb.set_trace() return None print('try {} times'.format(trails)) return scene def gen_caption(obj_name, obj_scale, ref_obj_name, ref_obj_scale, type='closer'): if type == 'closer': captions = [f'Move the {obj_name} and the {ref_obj_name} closer together.', f'Move the {obj_name} and the {ref_obj_name} closer.', f'Move the {obj_name} and the {ref_obj_name} closer to each other.', f'Move both objects closer', f'Move the two objects closer together.', f'Move the two objects closer to each other.'] elif type == 'further': captions = [f'Move the {obj_name} further away from the {ref_obj_name}.', f'Move the {obj_name} and the {ref_obj_name} further apart.', f'Move the {obj_name} and the {ref_obj_name} further away from each other.', f'Move both objects further apart', f'Move the two objects further away from each other.', f'Move the two objects further apart.'] elif type == 'swap': captions = [f'Swap the positions of the {obj_name} and the {ref_obj_name}.', f'Swap the positions of the {ref_obj_name} and the {obj_name}.', f'Swap the positions of the two objects.', f'Swap positions of both items.'] return random.choice(captions) # --- Common setups & resources print('Generate {} Sample'.format(FLAGS.generate_idx)) scene, rng, output_dir, scratch_dir = kb.setup(FLAGS) output_dir = output_dir / FLAGS.sub_outputdir simulator = PyBullet(scene, scratch_dir) renderer = Blender(scene, scratch_dir, samples_per_pixel=64) kubasic = kb.AssetSource.from_manifest(FLAGS.kubasic_assets) gso = kb.AssetSource.from_manifest(FLAGS.gso_assets) hdri_source = kb.AssetSource.from_manifest(FLAGS.hdri_assets) # --- Populate the scene # background HDRI train_backgrounds, test_backgrounds = hdri_source.get_test_split(fraction=0.) logging.info("Choosing one of the %d training backgrounds...", len(train_backgrounds)) hdri_id = rng.choice(train_backgrounds) background_hdri = hdri_source.create(asset_id=hdri_id) #assert isinstance(background_hdri, kb.Texture) logging.info("Using background %s", hdri_id) scene.metadata["background"] = hdri_id renderer._set_ambient_light_hdri(background_hdri.filename) # Dome dome = kubasic.create(asset_id="dome", name="dome", friction=FLAGS.floor_friction, restitution=FLAGS.floor_restitution, static=True, background=True) assert isinstance(dome, kb.FileBasedObject) scene += dome dome_blender = dome.linked_objects[renderer] texture_node = dome_blender.data.materials[0].node_tree.nodes["Image Texture"] texture_node.image = bpy.data.images.load(background_hdri.filename) def get_linear_camera_motion_start_end( movement_speed: float, inner_radius: float = 8., outer_radius: float = 12., z_offset: float = 0.1, ): """Sample a linear path which starts and ends within a half-sphere shell.""" while True: camera_start = np.array(kb.sample_point_in_half_sphere_shell(inner_radius, outer_radius, z_offset)) direction = rng.rand(3) - 0.5 movement = direction / np.linalg.norm(direction) * movement_speed camera_end = camera_start + movement if (inner_radius <= np.linalg.norm(camera_end) <= outer_radius and camera_end[2] > z_offset): return camera_start, camera_end # Camera logging.info("Setting up the Camera...") scene.camera = kb.PerspectiveCamera(focal_length=35., sensor_width=36) if FLAGS.camera == "fixed_random": # scene.camera.position = kb.sample_point_in_half_sphere_shell( # inner_radius=7., outer_radius=9., offset=4) scene.camera.position = (0, -10, 15) scene.camera.look_at((0, 0, 0)) elif FLAGS.camera == "linear_movement": camera_start, camera_end = get_linear_camera_motion_start_end( movement_speed=rng.uniform(low=0., high=FLAGS.max_camera_movement) ) # linearly interpolate the camera position between these two points # while keeping it focused on the center of the scene # we start one frame early and end one frame late to ensure that # forward and backward flow are still consistent for the last and first frames for frame in range(FLAGS.frame_start - 1, FLAGS.frame_end + 2): interp = ((frame - FLAGS.frame_start + 1) / (FLAGS.frame_end - FLAGS.frame_start + 3)) scene.camera.position = (interp * np.array(camera_start) + (1 - interp) * np.array(camera_end)) scene.camera.look_at((0, 0, 0)) scene.camera.keyframe_insert("position", frame) scene.camera.keyframe_insert("quaternion", frame) # Add random objects train_split, test_split = gso.get_test_split(fraction=0.) # if FLAGS.objects_split == "train": logging.info("Choosing one of the %d training objects...", len(train_split)) # active_split = train_split active_split = list(GSO_dict.keys()) # num_objects = rng.randint(FLAGS.min_num_objects, # FLAGS.max_num_objects+1) num_objects = 2 logging.info("Step 1: Randomly placing %d objects:", num_objects) object_state_save_dict = {} object_state_ref_dict = {} # not resample objects object_id_list = random.sample(active_split, num_objects+1) for i in range(num_objects): # object_id = rng.choice(active_split) object_id = object_id_list[i] obj = gso.create(asset_id=object_id) assert isinstance(obj, kb.FileBasedObject) scale = rng.uniform(FLAGS.smallest_scale, FLAGS.largest_scale) obj.scale = scale / np.max(obj.bounds[1] - obj.bounds[0]) obj_pos_z = - get_obj_z(obj.bounds, obj.scale) SPAWN_REGION_OBJ[0][2], SPAWN_REGION_OBJ[1][2] = obj_pos_z, obj_pos_z obj.position = rng.uniform(*SPAWN_REGION_OBJ) obj.metadata["scale"] = scale scene += obj move_until_no_overlap(obj, simulator, spawn_region=SPAWN_REGION_OBJ, rng=rng) # initialize velocity randomly but biased towards center # obj.velocity = (rng.uniform(*VELOCITY_RANGE) - # [obj.position[0], obj.position[1], 0]) # print(obj.position) obj.velocity = [0, 0, 0] logging.info(" Added %s at %s", obj.asset_id, obj.position) object_state_save_dict[i] = {'object_id': object_id, 'object_scale': obj.scale, 'object_quaternion': obj.quaternion, 'object_bounds': obj.bounds} object_state_ref_dict[i] = {'object': obj} ref_object = object_state_ref_dict[list(object_state_ref_dict.keys())[0]]['object'] ref_object_name = GSO_dict[ref_object.asset_id] ref_location = ref_object.position obj = object_state_ref_dict[list(object_state_ref_dict.keys())[1]]['object'] new_object_name = GSO_dict[obj.asset_id] # 1st print('Generate the first scene.') # object_id = rng.choice(active_split) # object_id = object_id_list[-1] # obj = gso.create(asset_id=object_id) # scale = rng.uniform(FLAGS.smallest_scale, FLAGS.largest_scale) # obj.scale = scale / np.max(obj.bounds[1] - obj.bounds[0]) # obj.metadata["scale"] = scale # new_object_name = GSO_dict[obj.asset_id] # print('Add new object {}'.format(new_object_name)) # scene = add_new_obj(scene, obj, ref_location1, ref_object, rng, max_trails=500) if scene is None: exit() frame = renderer.render_still() edits = ['close', 'swap'] edit = random.choice(edits) os.makedirs(output_dir/'{}'.format(FLAGS.generate_idx), exist_ok=True) kb.write_png(frame["rgba"], output_dir/"{}/image0.png".format(FLAGS.generate_idx)) caption_1 = gen_caption(new_object_name, obj.metadata["scale"], ref_object_name, ref_object.metadata["scale"], type="further" if edit == 'close' else 'swap') print(caption_1) # save meta ann object_state_save_dict[i+1] = {'object_id': object_id, 'object_scale': obj.scale, 'object_pos': obj.position, 'object_quaternion': obj.quaternion, 'object_bounds': obj.bounds} # import json # json.dump(object_state_save_dict, open(output_dir/'{}/meta_ann1.json'.format(FLAGS.generate_idx), 'w')) # np.save(output_dir/'{}/meta_ann1.npy'.format(FLAGS.generate_idx), object_state_save_dict) # renderer.save_state(output_dir/'{}/image1.blend'.format(FLAGS.generate_idx)) # 2nd print('Generate the second scene.') # delete the last object to generate the second frame # import pdb; pdb.set_trace() # scene.remove(obj) # scene= add_new_obj(scene, obj, ref_location2, ref_object, rng, max_trails=100) ref_obj_pos = ref_object.position ref_obj_z_up = get_obj_z_up(ref_object.bounds, ref_object.scale) logging.info(f'Object position: {obj.position}') obj1_pos = obj.position obj2_pos = ref_location if edit == 'close': direction = obj2_pos - obj1_pos factor1, factor2 = random.uniform(0.1, 0.3), random.uniform(0.1, 0.3) obj.position = obj1_pos + direction * factor1 ref_object.position = obj2_pos - direction * factor2 else: obj.position = obj2_pos ref_object.position = obj1_pos frame = renderer.render_still() kb.write_png(frame["rgba"], output_dir/"{}/image1.png".format(FLAGS.generate_idx)) caption_2 = gen_caption(new_object_name, obj.metadata["scale"], ref_object_name, ref_object.metadata["scale"], type="closer" if edit == 'close' else 'swap') print(caption_2) # save meta ann object_state_save_dict[i+1] = {'object_id': object_id, 'object_scale': obj.scale, 'object_pos': obj.position, 'object_quaternion': obj.quaternion, 'object_bounds': obj.bounds} # import json # json.dump(object_state_save_dict, open(output_dir/'{}/meta_ann2.json'.format(FLAGS.generate_idx), 'w')) # np.save(output_dir/'{}/meta_ann2.npy'.format(FLAGS.generate_idx), object_state_save_dict) # renderer.save_state(output_dir/'{}/image2.blend'.format(FLAGS.generate_idx)) # save json # local_ann = {'image0':"{}/image0.png".format(FLAGS.generate_idx), 'caption0':caption_1, # 'image1':"{}/image1.png".format(FLAGS.generate_idx), 'caption1':caption_2, # 'ann_path':"{}/ann.json".format(FLAGS.generate_idx), # 'obj_num':num_objects+1} # json.dump(local_ann, open("{}/{}/ann.json".format(str(output_dir), FLAGS.generate_idx), 'w')) # import pdb; pdb.set_trace() # if not os.path.exists("{}/global_ann.json".format(str(output_dir))): # json.dump([], open("{}/global_ann.json".format(str(output_dir)), 'w')) # with open("{}/global_ann.json".format(str(output_dir)), 'r') as f: # old_data = json.load(f) # old_data.append(local_ann) # with open("{}/global_ann.json".format(str(output_dir)), "w") as f: # json.dump(old_data, f) local_ann = [{ 'input': dataset_dir(DATASET_TYPE) + "{}/image0.png".format(FLAGS.generate_idx), 'output': dataset_dir(DATASET_TYPE) + "{}/image1.png".format(FLAGS.generate_idx), 'instruction': caption_2, }, { 'input': dataset_dir(DATASET_TYPE) + "{}/image1.png".format(FLAGS.generate_idx), 'output': dataset_dir(DATASET_TYPE) + "{}/image0.png".format(FLAGS.generate_idx), 'instruction': caption_1, } ] save_scene_instruction(f"{output_dir}/eq_kubric_{DATASET_TYPE}.json", local_ann, DATASET_TYPE, FLAGS.generate_idx) kb.done()