Uni3C/cam_render.py at main · alibaba-damo-academy/Uni3C · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
import argparse
import json
import os
import warnings

import einops
import numpy as np
import torch
import trimesh
from PIL import Image, ImageOps
from carvekit.ml.wrap.tracer_b7 import TracerUniversalB7
from diffusers.utils import export_to_video
from pytorch3d.renderer import PointsRasterizationSettings
from torchvision.transforms import ToTensor, ToPILImage

import third_party.depth_pro.depth_pro as depth_pro
from src.pointcloud import point_rendering
from src.utils import traj_map, points_padding, np_points_padding, set_initial_camera, build_cameras
from huggingface_hub import hf_hub_download

warnings.filterwarnings("ignore")


def main():
    torch.set_grad_enabled(False)
    # == parse configs ==
    parser = argparse.ArgumentParser()
    parser.add_argument("--reference_image", default=None, type=str, required=True,
                        help="the path of input image")
    parser.add_argument("--output_path", default="outputs/temp", type=str,
                        help="output folder's path")
    parser.add_argument("--traj_type", default="custom", type=str,
                        choices=["custom", "free1", "free2", "free3", "free4", "free5", "swing1", "swing2", "orbit"],
                        help="custom refers to a custom trajectory, while the others are pre-defined camera trajectories (see traj_map for details)")
    parser.add_argument("--nframe", default=81, type=int, help="Total number of frames")
    parser.add_argument("--d_r", default=1.0, type=float,
                        help="Camera distance, default is 1.0, range 0.25 to 2.5")
    parser.add_argument("--d_theta", default=0.0, type=float,
                        help="Vertical rotation, <0 up, >0 down, range -90 to 30; generally not recommended to angle too much downwards")
    parser.add_argument("--d_phi", default=0.0, type=float,
                        help="Horizontal rotation, <0 right, >0 left, supports 360 degrees; range -360 to 360")
    parser.add_argument("--x_offset", default=0.0, type=float,
                        help="Horizontal translation, <0 left, >0 right, range -0.5 to 0.5; depends on depth, excessive movement may cause artifacts")
    parser.add_argument("--y_offset", default=0.0, type=float,
                        help="Vertical translation, <0 up, >0 down, range -0.5 to 0.5; depends on depth, excessive movement may cause artifacts")
    parser.add_argument("--z_offset", default=0.0, type=float,
                        help="Forward and backward translation, <0 back, >0 forward, range -0.5 to 0.5 is ok; depends on depth, excessive movement may cause artifacts")
    parser.add_argument("--focal_length", default=1.0, type=float,
                        help="Focal length, range 0.25 to 2.5; changing focal length zooms in and out")
    parser.add_argument("--start_elevation", default=5.0, type=float,
                        help="Initial angle, no exceptions to change")

    args = parser.parse_args()
    device = "cuda"
    print("Init depth model")
    depth_model, depth_transform = depth_pro.create_model_and_transforms(device=device)
    depth_model = depth_model.eval()

    print("Init mask model")
    seg_net = TracerUniversalB7(device=device, batch_size=1,
                                model_path=hf_hub_download(repo_id="Carve/tracer_b7", filename="tracer_b7.pth")).eval()

    # == motion definition ==
    if args.traj_type == "custom":
        cam_traj, x_offset, y_offset, z_offset, d_theta, d_phi, d_r = \
            "free", args.x_offset, args.y_offset, args.z_offset, args.d_theta, args.d_phi, args.d_r
    else:
        cam_traj, x_offset, y_offset, z_offset, d_theta, d_phi, d_r = traj_map(args.traj_type)

    # load image
    image = Image.open(args.reference_image).convert("RGB")
    image = ImageOps.exif_transpose(image)
    w_origin, h_origin = image.size
    hw_list = [[480, 768], [512, 720], [608, 608], [720, 512], [768, 480]]
    hw_ratio_list = [h_ / w_ for [h_, w_] in hw_list]
    hw_ratio = h_origin / w_origin
    sub_hw_ratio = np.abs(np.array(hw_ratio_list) - hw_ratio)
    select_idx = np.argmin(sub_hw_ratio)
    height, width = hw_list[select_idx]
    print(f"Image: {args.reference_image.split('/')[-1]}, Resolution: {h_origin}x{w_origin}->{height}x{width}")
    image = image.resize((width, height), Image.Resampling.BICUBIC)
    validation_image = ToTensor()(image)[None]  # [1,c,h,w], 0~1

    os.makedirs(args.output_path, exist_ok=True)

    # inference depth
    with torch.no_grad():
        depth_image = np.array(image)
        depth_image = depth_transform(depth_image)
        prediction = depth_model.infer(depth_image, f_px=None)
        depth = prediction["depth"]  # Depth in [m].
        depth = depth[None, None]
        focallength_px = prediction["focallength_px"].item()  # Focal length in pixels.
        K = torch.tensor([[focallength_px, 0, width / 2],
                          [0, focallength_px, height / 2],
                          [0, 0, 1]], dtype=torch.float32)
        K_inv = K.inverse()
        intrinsic = K[None].repeat(args.nframe, 1, 1)

    # get pointcloud
    points2d = torch.stack(torch.meshgrid(torch.arange(width, dtype=torch.float32),
                                          torch.arange(height, dtype=torch.float32), indexing="xy"), -1)  # [h,w,2]
    points3d = points_padding(points2d).reshape(height * width, 3)  # [hw,3]
    points3d = (K_inv @ points3d.T * depth.reshape(1, height * width).cpu()).T
    colors = ((depth_image + 1) / 2 * 255).to(torch.uint8).permute(1, 2, 0).reshape(height * width, 3)
    points3d = points3d.cpu().numpy()
    colors = colors.cpu().numpy()

    # inference foreground mask
    with torch.no_grad():
        origin_w_, origin_h_ = image.size
        image_pil = image.resize((512, 512))
        fg_mask = seg_net([image_pil])[0]
        fg_mask = fg_mask.resize((origin_w_, origin_h_))
    fg_mask = np.array(fg_mask)
    fg_mask = fg_mask > 127.5
    fg_mask = torch.tensor(fg_mask)
    if fg_mask.float().mean() < 0.05:
        fg_mask[...] = True
    depth_avg = torch.median(depth[0, 0, fg_mask]).item()

    w2c_0, c2w_0 = set_initial_camera(args.start_elevation, depth_avg)

    # convert points3d to the world coordinate
    points3d = (c2w_0.numpy()[:3] @ np_points_padding(points3d).T).T
    pcd = trimesh.PointCloud(vertices=points3d, colors=colors)
    _ = pcd.export(f"{args.output_path}/pcd.ply")

    # build camera viewpoints according to d_theta，d_phi, d_r
    w2cs, c2ws, intrinsic = build_cameras(cam_traj=cam_traj,
                                          w2c_0=w2c_0,
                                          c2w_0=c2w_0,
                                          intrinsic=intrinsic,
                                          nframe=args.nframe,
                                          focal_length=args.focal_length,
                                          d_theta=d_theta,
                                          d_phi=d_phi,
                                          d_r=d_r,
                                          radius=depth_avg,
                                          x_offset=x_offset,
                                          y_offset=y_offset,
                                          z_offset=z_offset)

    # save camera infos
    w2cs_list = w2cs.cpu().numpy().tolist()
    camera_infos = {"intrinsic": K.cpu().numpy().tolist(), "extrinsic": w2cs_list, "height": height, "width": width}
    with open(f"{args.output_path}/cam_info.json", "w") as writer:
        json.dump(camera_infos, writer, indent=2)

    with torch.no_grad(), torch.amp.autocast("cuda", enabled=False):
        control_imgs, render_masks = point_rendering(K=intrinsic.float(),
                                                     w2cs=w2cs.float(),
                                                     depth=depth.float(),
                                                     image=validation_image.float() * 2 - 1,
                                                     raster_settings=PointsRasterizationSettings(image_size=(height, width),
                                                                                                 radius=0.008,
                                                                                                 points_per_pixel=8),
                                                     device=device,
                                                     background_color=[0, 0, 0],
                                                     sobel_threshold=0.35,
                                                     sam_mask=None)

    control_imgs = einops.rearrange(control_imgs, "(b f) c h w -> b c f h w", f=args.nframe)
    render_masks = einops.rearrange(render_masks, "(b f) c h w -> b c f h w", f=args.nframe)

    render_video = []
    mask_video = []
    control_imgs = control_imgs.to(torch.float32)
    for i in range(args.nframe):
        img = ToPILImage()((control_imgs[0][:, i] + 1) / 2)
        render_video.append(img)
        mask = ToPILImage()(render_masks[0][:, i])
        mask_video.append(mask)

    export_to_video(render_video, f"{args.output_path}/render.mp4", fps=16)
    export_to_video(mask_video, f"{args.output_path}/render_mask.mp4", fps=16)

    print("Rendering finished.")


if __name__ == "__main__":
    main()