File size: 14,405 Bytes
2913579 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 |
# 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.
#!/usr/bin/env python3
"""
fast3r_re10k_pose_eval_multi_gpu.py
1) Root set to /home/jianingy/research/fast3r/src via rootutils.
2) Splits RealEstate10K test folders between 2 GPUs in parallel (by default).
3) Loads & runs DUSt3R model on each GPU, aggregates final results.
4) Allows evaluating only a subset of scene folders (via --subset_file).
5) Correctly parses RealEstate10K extrinsics as c2w.
"""
import os
import glob
import random
import time
import copy
import math
import json
from collections import defaultdict
from multiprocessing import get_context
import numpy as np
from numpy.linalg import inv
import torch
import hydra
import open3d as o3d
import trimesh
from omegaconf import OmegaConf
from PIL import Image
from tqdm import tqdm
# For CLI
import argparse
# Attempt pretty console table with 'rich'
try:
from rich.table import Table
from rich.console import Console
RICH_AVAILABLE = True
except ImportError:
RICH_AVAILABLE = False
# Rootutils: set project root => /home/jianingy/research/fast3r/src
import rootutils
rootutils.setup_root(
"/home/jianingy/research/fast3r/src",
indicator=".project-root", # or remove if not needed
pythonpath=True
)
# Project-specific imports
from fast3r.models.multiview_dust3r_module import MultiViewDUSt3RLitModule
from fast3r.dust3r.inference_multiview import inference
from fast3r.dust3r.model import FlashDUSt3R
from fast3r.dust3r.utils.image import rgb, imread_cv2
from fast3r.dust3r.datasets.utils.transforms import ImgNorm
import fast3r.dust3r.datasets.utils.cropping as cropping
from lightning.pytorch.utilities.deepspeed import convert_zero_checkpoint_to_fp32_state_dict
########################
# 1) Utility: fix old references in config
########################
def replace_dust3r_in_config(cfg):
for key, value in cfg.items():
if isinstance(value, dict):
replace_dust3r_in_config(value)
elif isinstance(value, str):
if "dust3r." in value and "fast3r.dust3r." not in value:
cfg[key] = value.replace("dust3r.", "fast3r.dust3r.")
return cfg
########################
# 2) Crop & Resize
########################
def crop_resize_if_necessary(
image,
intrinsics_3x3,
target_resolution=(512, 288),
rng=None,
info=None
):
"""Crop around principal point + downscale => (512×288) or (288×512)."""
if isinstance(image, np.ndarray):
image = Image.fromarray(image)
W_org, H_org = image.size
cx, cy = int(round(intrinsics_3x3[0, 2])), int(round(intrinsics_3x3[1, 2]))
min_margin_x = min(cx, W_org - cx)
min_margin_y = min(cy, H_org - cy)
left = cx - min_margin_x
top = cy - min_margin_y
right = cx + min_margin_x
bottom = cy + min_margin_y
crop_bbox = (left, top, right, bottom)
image_c, _, intrinsics_c = cropping.crop_image_depthmap(
image,
None,
intrinsics_3x3,
crop_bbox
)
# Check orientation
W_c, H_c = image_c.size
if H_c > W_c:
# swap if portrait
target_resolution = (target_resolution[1], target_resolution[0])
# Downscale
image_rs, _, intrinsics_rs = cropping.rescale_image_depthmap(
image_c, None, intrinsics_c, np.array(target_resolution)
)
intrinsics2 = cropping.camera_matrix_of_crop(
intrinsics_rs, image_rs.size, target_resolution, offset_factor=0.5
)
final_bbox = cropping.bbox_from_intrinsics_in_out(
intrinsics_rs, intrinsics2, target_resolution
)
image_out, _, intrinsics_out = cropping.crop_image_depthmap(
image_rs, None, intrinsics_rs, final_bbox
)
return image_out, intrinsics_out
########################
# 3) Worker function: processes a subset of folders on a given GPU
########################
def process_folders(args):
"""
args: tuple(
video_folders, device_id,
re10k_video_root, re10k_txt_root,
checkpoint_path, output_folder
)
Each worker loads the model on 'cuda:device_id' and runs inference.
Returns a list of dicts with final metrics for each folder.
"""
(video_folders, device_id,
re10k_video_root, re10k_txt_root,
checkpoint_path, output_folder) = args
device = torch.device(f"cuda:{device_id}" if torch.cuda.is_available() else "cpu")
print(f"[Process GPU {device_id}] => device = {device}")
# 1) Load model
print(f"[Process GPU {device_id}] Loading checkpoint: {checkpoint_path}")
cfg_path = os.path.join(os.path.dirname(checkpoint_path), "../.hydra/config.yaml")
cfg = OmegaConf.load(cfg_path)
cfg.model.net = replace_dust3r_in_config(cfg.model.net)
if "encoder_args" in cfg.model.net:
cfg.model.net.encoder_args.patch_embed_cls = "PatchEmbedDust3R"
cfg.model.net.head_args.landscape_only = False
else:
cfg.model.net.patch_embed_cls = "PatchEmbedDust3R"
cfg.model.net.landscape_only = False
cfg.model.net.decoder_args.random_image_idx_embedding = True
cfg.model.net.decoder_args.attn_bias_for_inference_enabled = False
lit_module = hydra.utils.instantiate(
cfg.model, train_criterion=None, validation_criterion=None
)
lit_module = MultiViewDUSt3RLitModule.load_from_checkpoint(
checkpoint_path=checkpoint_path,
net=lit_module.net,
train_criterion=lit_module.train_criterion,
validation_criterion=lit_module.validation_criterion
)
lit_module.eval()
model = lit_module.net.to(device)
# Optionally compile
# model = torch.compile(model)
results_list = []
# 2) For each folder => sample frames => run inference => evaluate poses
for vid_folder in tqdm(video_folders, desc=f"[GPU {device_id}] Evaluate"):
folder_path = os.path.join(re10k_video_root, vid_folder)
if not os.path.isdir(folder_path):
continue
txt_path = os.path.join(re10k_txt_root, vid_folder + ".txt")
if not os.path.exists(txt_path):
continue
with open(txt_path, "r") as f:
txt_lines = f.read().strip().split("\n")
if len(txt_lines) <= 1:
continue
txt_lines = txt_lines[1:] # skip first line (URL)
lines_map = {}
for line in txt_lines:
parts = line.strip().split()
# Expect at least 19 columns
if len(parts) < 19:
continue
frame_id = parts[0]
lines_map[frame_id] = parts
frame_files = sorted(glob.glob(os.path.join(folder_path, "*.jpg")))
if len(frame_files) < 2:
continue
# Sample up to 10 frames per folder
n_to_sample = min(10, len(frame_files))
sampled_frames = random.sample(frame_files, n_to_sample)
sampled_frames.sort()
selected_views = []
for frame_path in sampled_frames:
basename = os.path.splitext(os.path.basename(frame_path))[0]
if basename not in lines_map:
continue
columns = lines_map[basename]
# parse fx, fy, cx, cy
fx = float(columns[1])
fy = float(columns[2])
cx = float(columns[3])
cy = float(columns[4])
# parse extrinsic row-major 3×4 => build 4×4, then invert to get c2w
extrinsic_val = [float(v) for v in columns[7:19]] # 12 floats
extrinsic = np.array(extrinsic_val, dtype=np.float64).reshape(3, 4)
# Build 4x4
pose_4x4 = np.eye(4, dtype=np.float32)
pose_4x4[:3, :3] = extrinsic[:3, :3]
pose_4x4[:3, 3] = extrinsic[:3, 3] # translation in last column
poses_c2w_gt = inv(pose_4x4).astype(np.float32)
# Load image
img_rgb = imread_cv2(frame_path)
if img_rgb is None:
continue
H_org, W_org = img_rgb.shape[:2]
# Build K
K_3x3 = np.array([
[fx * W_org, 0.0, cx * W_org],
[0.0, fy * H_org, cy * H_org],
[0.0, 0.0, 1.0]
], dtype=np.float32)
# Crop+resize => final_img_pil, final_intrinsics
pil_img = Image.fromarray(img_rgb)
final_img_pil, final_intrinsics_3x3 = crop_resize_if_necessary(
pil_img, K_3x3, target_resolution=(512, 288)
)
tensor_chw = ImgNorm(final_img_pil)
# Put data on GPU
view_dict = {
"img": tensor_chw.unsqueeze(0).to(device), # (1,3,H,W)
"camera_pose": torch.from_numpy(poses_c2w_gt).unsqueeze(0).to(device), # (1,4,4)
"camera_intrinsics": torch.from_numpy(final_intrinsics_3x3).unsqueeze(0).to(device), # (1,3,3)
"dataset": ["RealEstate10K"],
"true_shape": torch.tensor([[final_img_pil.size[1], final_img_pil.size[0]]], device=device),
}
selected_views.append(view_dict)
if len(selected_views) < 2:
continue
output = inference(
selected_views,
model=model,
device=device,
dtype=torch.float32,
verbose=False,
profiling=False
)
# Evaluate camera poses
pose_results = lit_module.evaluate_camera_poses(
views=output["views"],
preds=output["preds"],
niter_PnP=100,
focal_length_estimation_method='first_view_from_global_head'
)
if len(pose_results) > 0:
metrics_dict = pose_results[0]
metrics_dict["video_name"] = vid_folder
# Save result to file
out_path = os.path.join(output_folder, f"{vid_folder}.txt")
with open(out_path, "w") as ff:
ff.write(str(metrics_dict))
results_list.append(metrics_dict)
return results_list
########################
# 4) The main: splits data for 2 GPUs, spawns processes, aggregates final metrics
########################
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--subset_file", type=str, default=None,
help="Optional path to a text file with subset folder names to evaluate")
parser.add_argument("--gpu_count", type=int, default=2, help="Number of GPUs to use")
args = parser.parse_args()
# Setup paths
re10k_video_root = "/data/jianingy/RealEstate10K/videos/test"
re10k_txt_root = "/data/jianingy/RealEstate10K/test"
output_folder = "/home/jianingy/research/fast3r/notebooks/RealEstate10K_eval"
os.makedirs(output_folder, exist_ok=True)
# Checkpoint
checkpoint_dir = "/data/jianingy/dust3r_data/fast3r_checkpoints/super_long_training_5175604"
possible_dir = os.path.join(checkpoint_dir, "checkpoints", "last.ckpt")
if os.path.isdir(possible_dir):
# Convert zero checkpoint
ckpt_agg = os.path.join(checkpoint_dir, "checkpoints", "last_aggregated.ckpt")
if not os.path.exists(ckpt_agg):
convert_zero_checkpoint_to_fp32_state_dict(possible_dir, ckpt_agg, tag=None)
CKPT_PATH = ckpt_agg
else:
CKPT_PATH = os.path.join(checkpoint_dir, "checkpoints", "last.ckpt")
# All video folders
all_folders = sorted(os.listdir(re10k_video_root))
all_folders = [f for f in all_folders if os.path.isdir(os.path.join(re10k_video_root, f))]
# If user specified a subset file, only keep those folders
if args.subset_file and os.path.exists(args.subset_file):
with open(args.subset_file, "r") as f:
subset_scenes = set(line.strip() for line in f if line.strip())
all_folders = [fd for fd in all_folders if fd in subset_scenes]
# If no folders remain, just exit
if not all_folders:
print("No matching folders found. Exiting.")
return
# Split in 'args.gpu_count' parts. By default =2
n_gpus = args.gpu_count
chunk_size = math.ceil(len(all_folders) / n_gpus)
subfolders_list = []
for i in range(n_gpus):
start_i = i * chunk_size
end_i = start_i + chunk_size
subfolders_list.append(all_folders[start_i:end_i])
# We'll run n_gpus processes in parallel
tasks = []
for i in range(n_gpus):
if subfolders_list[i]:
tasks.append((subfolders_list[i], i, re10k_video_root, re10k_txt_root, CKPT_PATH, output_folder))
ctx = get_context("spawn") # or "fork" if on Linux
pool = ctx.Pool(processes=len(tasks))
async_results = []
for arg_tuple in tasks:
ar = pool.apply_async(process_folders, (arg_tuple,))
async_results.append(ar)
# Collect
all_results = []
for ar in async_results:
subset_res = ar.get() # each is a list of metrics
all_results.extend(subset_res)
pool.close()
pool.join()
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Aggregate
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
aggregator = defaultdict(list)
# Typical keys from evaluate_camera_poses
metric_keys = ["RRA_at_5","RRA_at_15","RRA_at_30","RTA_at_5","RTA_at_15","RTA_at_30","mAA_30"]
for res in all_results:
for k in metric_keys:
if k in res:
aggregator[k].append(float(res[k]))
final_means = {}
for k in metric_keys:
vals = aggregator.get(k, [])
if vals:
final_means[k] = sum(vals)/len(vals)
else:
final_means[k] = float("nan")
# Print summary
if RICH_AVAILABLE:
console = Console()
table = Table(title="RealEstate10K Pose Metrics Summary (Multi-GPU aggregated)")
table.add_column("Metric", justify="left")
table.add_column("Mean Value", justify="right")
for k in metric_keys:
val = final_means[k]
table.add_row(k, f"{val:.4f}")
console.print(table)
else:
print("Final Means for RealEstate10K Pose:")
for k in metric_keys:
print(f"{k}: {final_means[k]:.4f}")
print(f"[Main] Done! Processed {len(all_results)} video folders total.")
if __name__ == "__main__":
main()
|