import sys sys.path.append('core') import os import torch import numpy as np from…
- User
import sys sys.path.append('core') import os import torch import numpy as np from pathlib import Path from osgeo import gdal # 新增GDAL from core.igev_stereo import IGEVStereo from core.utils.utils import InputPadder import argparse from tqdm import tqdm import imageio DEVICE = 'cuda' #from feature_match.eloftr.match_image import get_init_disp #from eloftr.match_image import get_init_disp,init_matcher from PIL import Image import cv2
取外扩 50 个像素
def load_image_block(dataset, x_offset, y_offset, x_size, y_size): """从GDAL数据集中读取图像块""" band = dataset.GetRasterBand(1) img_block = band.ReadAsArray(x_offset, y_offset, x_size, y_size).astype(np.uint8) # 读取指定位置的图像块 if len(img_block.shape) == 2: # 单通道转三通道 img_block = np.stack([img_block] * 3, axis=-1) img_block = torch.from_numpy(img_block).permute(2, 0, 1).float() # 转换为 PyTorch 张量 return img_block[None].to(DEVICE)
def get_disp_img_pyramid(model, image1_path, image2_path,disp_init, output_directory): # 使用 GDAL 打开图像 dataset1 = gdal.Open(image1_path) dataset2 = gdal.Open(image2_path) # 获取图像大小 width = dataset1.RasterXSize height = dataset1.RasterYSize disp_map = np.zeros((height, width), dtype=np.float32) weight_map = np.zeros((height, width), dtype=np.float32) block_sizes = 512 #128 #512 #128
w_block_number = width // block_sizes + 1 h_block_number = height // block_sizes + 1 for i in tqdm(range(h_block_number)): for j in range(w_block_number): #计算左图原始位置 l_w_start = j*block_sizes l_w_end = min(width,l_w_start+block_sizes) l_h_start = i*block_sizes l_h_end = min(height,l_h_start+block_sizes) #计算右图的原始位置 r_w_start = j*block_sizes r_w_end = min(width,r_w_start+block_sizes) r_h_start = i*block_sizes r_h_end = min(height,r_h_start+block_sizes) w_expand_pixels = 128 #512 #320 h_expand_pixels = 32 # 原始 block_init_disp 计算 block_init_disp = np.nanmean(disp_init[l_h_start:l_h_end, l_w_start:l_w_end]) if abs(block_init_disp)>256: w_expand_pixels += int(abs(block_init_disp)/2) h_expand_pixels += 64 # 如果整个区域都是 NaN,可以选择默认值,例如 0 if np.isnan(block_init_disp): block_init_disp = 0 else: block_init_disp = int(block_init_disp) #计算右图初始位置 skew_r_w_start = r_w_start-block_init_disp skew_r_w_end = r_w_end-block_init_disp skew_disp = 0 #如果超出范围,我就不管了 if skew_r_w_start>0 and skew_r_w_end < width: r_w_start = skew_r_w_start r_w_end = skew_r_w_end skew_disp = block_init_disp l_w_start_adjusted = max(0, l_w_start - w_expand_pixels) l_w_end_adjusted = min(width, l_w_end + w_expand_pixels) l_h_start_adjusted = max(0, l_h_start - h_expand_pixels) l_h_end_adjusted = min(height, l_h_end + h_expand_pixels) r_w_start_adjusted = max(0, r_w_start - w_expand_pixels) r_w_end_adjusted = min(width, r_w_end + w_expand_pixels) r_h_start_adjusted = max(0, r_h_start - h_expand_pixels) r_h_end_adjusted = min(height, r_h_end + h_expand_pixels) #根据左图的外扩计算右图然后也得根据右图能不能外扩取计算左图 map_left_expand = min(l_w_start-l_w_start_adjusted,r_w_start-r_w_start_adjusted) map_right_expand = min(l_w_end_adjusted - l_w_end,r_w_end_adjusted-r_w_end) map_up_expand = min(l_h_start-l_h_start_adjusted,r_h_start-r_h_start_adjusted) map_down_expand = min(l_h_end_adjusted - l_h_end,r_h_end_adjusted-r_h_end) l_w_start_adjusted = l_w_start-map_left_expand l_w_end_adjusted = l_w_end+map_right_expand l_h_start_adjusted = l_h_start-map_up_expand l_h_end_adjusted = l_h_end+map_down_expand r_w_start_adjusted = r_w_start - map_left_expand r_w_end_adjusted = r_w_end+map_right_expand r_h_start_adjusted = r_h_start - map_up_expand r_h_end_adjusted = r_h_end+map_down_expand image1_patch = load_image_block(dataset1, l_w_start_adjusted, l_h_start_adjusted, l_w_end_adjusted - l_w_start_adjusted, l_h_end_adjusted-l_h_start_adjusted) image2_patch = load_image_block(dataset2, r_w_start_adjusted, r_h_start_adjusted, r_w_end_adjusted - r_w_start_adjusted, r_h_end_adjusted-r_h_start_adjusted) padder = InputPadder(image1_patch.shape, divis_by=32) image1_patch, image2_patch = padder.pad(image1_patch, image2_patch) disp_patch = model(image1_patch, image2_patch, iters=12, test_mode=True) disp_patch = disp_patch.detach().cpu() disp_patch = padder.unpad(disp_patch).numpy()[0,0] #根据前面的外扩,再取出非外扩的值 up_expand = l_h_start - l_h_start_adjusted down_expand = l_h_end_adjusted - l_h_end left_expand = max(0,l_w_start - l_w_start_adjusted) right_expand = max(0,l_w_end_adjusted - l_w_end) # 去除外扩部分 disp_path_non_expanded = None if(right_expand==0): if(down_expand==0): disp_path_non_expanded = disp_patch[ up_expand:, left_expand: ] l_w_end = l_w_end_adjusted else: disp_path_non_expanded = disp_patch[ up_expand:-down_expand, left_expand: ] l_w_end = l_w_end_adjusted else: if(down_expand==0): disp_path_non_expanded = disp_patch[ up_expand:, left_expand:-right_expand ] else: disp_path_non_expanded = disp_patch[ up_expand:-down_expand, left_expand:-right_expand ] disp_map[l_h_start:l_h_end, l_w_start:l_w_end] += (disp_path_non_expanded+skew_disp) print(f"{l_h_start}_{l_h_end}_{l_w_start}_{l_w_end}") file_stem = image1_path.split('\\')[-2] filename = os.path.join(output_directory, f"{file_stem}_18.tif") imageio.imwrite(filename, disp_map.squeeze()) # 当数据集不再需要时关闭它们 dataset1 = None dataset2 = Nonedef load_image(imfile): # 打开图像文件 img = None img = Image.open(imfile)
# 计算原始图像的宽度和高度 original_width, original_height = img.size scale_size = 3096 # 确定缩放比例 scale = min(scale_size / original_width, scale_size / original_height) # 如果需要缩放,则进行缩放 if scale < 1: # 计算新的尺寸 new_width = int(original_width * scale) new_height = int(original_height * scale) # 等比例缩放图像 img = img.resize((new_width, new_height)) # 将图像转换为 numpy 数组 img = np.array(img).astype(np.uint8) img = np.stack([img]*3, axis=-1) img = torch.from_numpy(img).permute(2, 0, 1).float() return img[None].to(DEVICE),scaledef get_init_disp(left,right,model): test = True if test: # "D:\jx\ai\stereo_match_dl\IGEV-main\IGEV-main\IGEV-Stereo\core\demo-output\test_patch\SV1-02_20190416_L1B0000976744_8041900187280009_01-PAN2_epi_right_half.tif_1.tif" disp_resized_path = r"D:\jx\ai\stereo_match_dl\IGEV-main\IGEV-main\IGEV-Stereo\core\demo-output\test_patch\result_resize.tif" disp_resized = imageio.imread(disp_resized_path) return disp_resized else:
image1,scale = load_image(left) image2,scale = load_image(right) padder = InputPadder(image1.shape, divis_by=32) image1, image2 = padder.pad(image1, image2) disp = model(image1, image2, iters=args.valid_iters, test_mode=True) disp = disp.detach().cpu().numpy() disp = padder.unpad(disp) #再根据scale还原回去 disp_resized = cv2.resize(disp.squeeze(), (int(disp.shape[3] / scale), int(disp.shape[2] / scale)), interpolation=cv2.INTER_LINEAR)/scale return disp_resizedif name == 'main': parser = argparse.ArgumentParser() parser.add_argument('--restore_ckpt', help="restore checkpoint", default=r"D:\jx\ai\stereo_match_dl\IGEV-main\IGEV-main\IGEV-Stereo\ckpt\41481_epoch_igev-stereo.pth") # parser.add_argument('-l', '--left_imgs', help="path to all first (left) frames", default=r"D:\jx\ai\stereo_match_dl\test_image\bak\SV1-02_20190416_L1B0000976744_8041900187280009_01-PAN2_epi.tif") # parser.add_argument('-r', '--right_imgs', help="path to all second (right) frames", default=r"D:\jx\ai\stereo_match_dl\test_image\bak\SV1-02_20190416_L1B0000976744_8041900187280009_01-PAN1_epi.tif") parser.add_argument('-l', '--left_imgs', help="path to all first (left) frames", default=r"D:\jx\ai\stereo_match_dl\test_image\bak\SV1_PAN1_down_half_resized.tif") parser.add_argument('-r', '--right_imgs', help="path to all second (right) frames", default=r"D:\jx\ai\stereo_match_dl\test_image\bak\SV1_PAN2_down_half_resized.tif")
# parser.add_argument('-l', '--left_imgs', help="path to all first (left) frames", default=r"D:\jx\ai\stereo_match_dl\test_image\right\3_6_right.tif") # parser.add_argument('-r', '--right_imgs', help="path to all second (right) frames", default=r"D:\jx\ai\stereo_match_dl\test_image\left\3_6_left.tif") #parser.add_argument('-l', '--left_imgs', help="path to all first (left) frames", default=r"D:\jx\ai\stereo_match_dl\train_image\experimental data\with ground truth\test\left\HY_left_0.tiff") #parser.add_argument('-r', '--right_imgs', help="path to all second (right) frames", default=r"D:\jx\ai\stereo_match_dl\train_image\experimental data\with ground truth\test\right\HY_right_0.tiff") parser.add_argument('--output_directory', help="directory to save output", default=r"D:\jx\ai\stereo_match_dl\IGEV-main\IGEV-main\IGEV-Stereo\core\demo-output\test_patch") parser.add_argument('--mixed_precision', action='store_true', help='use mixed precision') # Architecture choices parser.add_argument('--hidden_dims', nargs='+', type=int, default=[128]*3, help="hidden state and context dimensions") parser.add_argument('--corr_implementation', choices=["reg", "alt", "reg_cuda", "alt_cuda"], default="reg", help="correlation volume implementation") parser.add_argument('--shared_backbone', action='store_true', help="use a single backbone for the context and feature encoders") parser.add_argument('--corr_levels', type=int, default=2, help="number of levels in the correlation pyramid") parser.add_argument('--corr_radius', type=int, default=4, help="width of the correlation pyramid") parser.add_argument('--n_downsample', type=int, default=2, help="resolution of the disparity field (1/2^K)") parser.add_argument('--slow_fast_gru', action='store_true', help="iterate the low-res GRUs more frequently") parser.add_argument('--n_gru_layers', type=int, default=3, help="number of hidden GRU levels") parser.add_argument('--max_disp', type=int, default=192, help="max disp of geometry encoding volume") args = parser.parse_args() Path(args.output_directory).mkdir(exist_ok=True, parents=True) # 初始化模型 model = torch.nn.DataParallel(IGEVStereo(args), device_ids=[0]) model.load_state_dict(torch.load(args.restore_ckpt)) model = model.module model.to(DEVICE) model.eval() result = get_init_disp(args.left_imgs,args.right_imgs,model) get_disp_img_pyramid(model, args.left_imgs,args.right_imgs, result ,args.output_directory)我这个,拼缝,有点明显,你帮我修改一下,加一个权重矩阵?
- Assistant
为了减少拼缝处的明显边界,可以在块之间引入权重矩阵(weight matrix)。这种方法通常通过给每个块分配一个平滑的权重,比如在块的中心区域权重较高,边缘区域权重逐渐减小,从而在重叠区域实现平滑过渡,减少拼缝的可见性。
下面是对您现有代码的修改建议,主要在
get_disp_img_pyramid函数中加入权重矩阵的处理:-
定义权重矩阵: 使用线性或高斯权重,确保中心区域权重更高,边缘区域逐渐减小。例如,使用线性渐变权重。
-
应用权重矩阵: 在将每个块的视差值(disparity)累加到
disp_map时,先乘以权重矩阵。同时,将权重矩阵累加到weight_map。 -
最终归一化: 在所有块处理完成后,将
disp_map除以weight_map,得到加权平均后的视差图。
以下是具体的代码修改示例:
import sys sys.path.append('core') import os import torch import numpy as np from pathlib import Path from osgeo import gdal # 新增GDAL from core.igev_stereo import IGEVStereo from core.utils.utils import InputPadder import argparse from tqdm import tqdm import imageio DEVICE = 'cuda' #from feature_match.eloftr.match_image import get_init_disp #from eloftr.match_image import get_init_disp,init_matcher from PIL import Image import cv2 def load_image_block(dataset, x_offset, y_offset, x_size, y_size): """从GDAL数据集中读取图像块""" band = dataset.GetRasterBand(1) img_block = band.ReadAsArray(x_offset, y_offset, x_size, y_size).astype(np.uint8) # 读取指定位置的图像块 if len(img_block.shape) == 2: # 单通道转三通道 img_block = np.stack([img_block] * 3, axis=-1) img_block = torch.from_numpy(img_block).permute(2, 0, 1).float() # 转换为 PyTorch 张量 return img_block[None].to(DEVICE) def create_weight_mask(height, width): """ 创建一个线性权重矩阵,中心权重最大,边缘权重最小。 这里使用简单的线性渐变,也可以使用高斯权重。 """ # 创建水平和垂直方向的线性权重 weight_x = np.linspace(0, 1, width) weight_y = np.linspace(0, 1, height) weight = np.outer(weight_y, weight_x) # 归一化权重到 [0,1] weight = weight / np.max(weight) return weight.astype(np.float32) def get_disp_img_pyramid(model, image1_path, image2_path, disp_init, output_directory): # 使用 GDAL 打开图像 dataset1 = gdal.Open(image1_path) dataset2 = gdal.Open(image2_path) # 获取图像大小 width = dataset1.RasterXSize height = dataset1.RasterYSize disp_map = np.zeros((height, width), dtype=np.float32) weight_map = np.zeros((height, width), dtype=np.float32) block_sizes = 512 #128 #512 #128 w_block_number = width // block_sizes + 1 h_block_number = height // block_sizes + 1 for i in tqdm(range(h_block_number)): for j in range(w_block_number): # 计算左图原始位置 l_w_start = j * block_sizes l_w_end = min(width, l_w_start + block_sizes) l_h_start = i * block_sizes l_h_end = min(height, l_h_start + block_sizes) # 计算右图的原始位置 r_w_start = j * block_sizes r_w_end = min(width, r_w_start + block_sizes) r_h_start = i * block_sizes r_h_end = min(height, r_h_start + block_sizes) w_expand_pixels = 128 #512 #320 h_expand_pixels = 32 # 原始 block_init_disp 计算 block_init_disp = np.nanmean(disp_init[l_h_start:l_h_end, l_w_start:l_w_end]) if abs(block_init_disp) > 256: w_expand_pixels += int(abs(block_init_disp) / 2) h_expand_pixels += 64 # 如果整个区域都是 NaN,可以选择默认值,例如 0 if np.isnan(block_init_disp): block_init_disp = 0 else: block_init_disp = int(block_init_disp) # 计算右图初始位置 skew_r_w_start = r_w_start - block_init_disp skew_r_w_end = r_w_end - block_init_disp skew_disp = 0 # 如果超出范围,不进行偏移 if skew_r_w_start > 0 and skew_r_w_end < width: r_w_start = skew_r_w_start r_w_end = skew_r_w_end skew_disp = block_init_disp l_w_start_adjusted = max(0, l_w_start - w_expand_pixels) l_w_end_adjusted = min(width, l_w_end + w_expand_pixels) l_h_start_adjusted = max(0, l_h_start - h_expand_pixels) l_h_end_adjusted = min(height, l_h_end + h_expand_pixels) r_w_start_adjusted = max(0, r_w_start - w_expand_pixels) r_w_end_adjusted = min(width, r_w_end + w_expand_pixels) r_h_start_adjusted = max(0, r_h_start - h_expand_pixels) r_h_end_adjusted = min(height, r_h_end + h_expand_pixels) # 根据左图的外扩计算右图然后也得根据右图能不能外扩取计算左图 map_left_expand = min(l_w_start - l_w_start_adjusted, r_w_start - r_w_start_adjusted) map_right_expand = min(l_w_end_adjusted - l_w_end, r_w_end_adjusted - r_w_end) map_up_expand = min(l_h_start - l_h_start_adjusted, r_h_start - r_h_start_adjusted) map_down_expand = min(l_h_end_adjusted - l_h_end, r_h_end_adjusted - r_h_end) l_w_start_adjusted = l_w_start - map_left_expand l_w_end_adjusted = l_w_end + map_right_expand l_h_start_adjusted = l_h_start - map_up_expand l_h_end_adjusted = l_h_end + map_down_expand r_w_start_adjusted = r_w_start - map_left_expand r_w_end_adjusted = r_w_end + map_right_expand r_h_start_adjusted = r_h_start - map_up_expand r_h_end_adjusted = r_h_end + map_down_expand image1_patch = load_image_block(dataset1, l_w_start_adjusted, l_h_start_adjusted, l_w_end_adjusted - l_w_start_adjusted, l_h_end_adjusted - l_h_start_adjusted) image2_patch = load_image_block(dataset2, r_w_start_adjusted, r_h_start_adjusted, r_w_end_adjusted - r_w_start_adjusted, r_h_end_adjusted - r_h_start_adjusted) padder = InputPadder(image1_patch.shape, divis_by=32) image1_patch, image2_patch = padder.pad(image1_patch, image2_patch) with torch.no_grad(): disp_patch = model(image1_patch, image2_patch, iters=12, test_mode=True) disp_patch = disp_patch.detach().cpu() disp_patch = padder.unpad(disp_patch).numpy()[0,0] # 根据前面的外扩,再取出非外扩的值 up_expand = l_h_start - l_h_start_adjusted down_expand = l_h_end_adjusted - l_h_end left_expand = max(0, l_w_start - l_w_start_adjusted) right_expand = max(0, l_w_end_adjusted - l_w_end) # 去除外扩部分 if right_expand == 0: if down_expand == 0: disp_patch_non_expanded = disp_patch[ up_expand:, left_expand: ] else: disp_patch_non_expanded = disp_patch[ up_expand:-down_expand, left_expand: ] else: if down_expand == 0: disp_patch_non_expanded = disp_patch[ up_expand:, left_expand:-right_expand ] else: disp_patch_non_expanded = disp_patch[ up_expand:-down_expand, left_expand:-right_expand ] # 创建与 disp_patch_non_expanded 尺寸匹配的权重矩阵 mask = create_weight_mask(disp_patch_non_expanded.shape[0], disp_patch_non_expanded.shape[1]) # 将权重应用于 disp_patch_non_expanded disp_map[l_h_start:l_h_end, l_w_start:l_w_end] += disp_patch_non_expanded * mask weight_map[l_h_start:l_h_end, l_w_start:l_w_end] += mask + skew_disp print(f"{l_h_start}_{l_h_end}_{l_w_start}_{l_w_end}") # 归一化 disparity map non_zero_mask = weight_map > 0 disp_map[non_zero_mask] = disp_map[non_zero_mask] / weight_map[non_zero_mask] # 将无效区域设为0或其他默认值 disp_map[~non_zero_mask] = 0 file_stem = os.path.basename(os.path.dirname(image1_path)) filename = os.path.join(output_directory, f"{file_stem}_18.tif") imageio.imwrite(filename, disp_map.astype(np.float32)) # 当数据集不再需要时关闭它们 dataset1 = None dataset2 = None def load_image(imfile): # 打开图像文件 img = None img = Image.open(imfile) # 计算原始图像的宽度和高度 original_width, original_height = img.size scale_size = 3096 # 确定缩放比例 scale = min(scale_size / original_width, scale_size / original_height) # 如果需要缩放,则进行缩放 if scale < 1: # 计算新的尺寸 new_width = int(original_width * scale) new_height = int(original_height * scale) # 等比例缩放图像 img = img.resize((new_width, new_height)) # 将图像转换为 numpy 数组 img = np.array(img).astype(np.uint8) img = np.stack([img]*3, axis=-1) img = torch.from_numpy(img).permute(2, 0, 1).float() return img[None].to(DEVICE), scale def get_init_disp(left, right, model): test = True if test: # "D:\jx\ai\stereo_match_dl\IGEV-main\IGEV-main\IGEV-Stereo\core\demo-output\test_patch\SV1-02_20190416_L1B0000976744_8041900187280009_01-PAN2_epi_right_half.tif_1.tif" disp_resized_path = r"D:\jx\ai\stereo_match_dl\IGEV-main\IGEV-main\IGEV-Stereo\core\demo-output\test_patch\result_resize.tif" disp_resized = imageio.imread(disp_resized_path) return disp_resized else: image1, scale = load_image(left) image2, scale = load_image(right) padder = InputPadder(image1.shape, divis_by=32) image1, image2 = padder.pad(image1, image2) disp = model(image1, image2, iters=args.valid_iters, test_mode=True) disp = disp.detach().cpu().numpy() disp = padder.unpad(disp) # 再根据scale还原回去 disp_resized = cv2.resize(disp.squeeze(), (int(disp.shape[3] / scale), int(disp.shape[2] / scale)), interpolation=cv2.INTER_LINEAR) / scale return disp_resized if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--restore_ckpt', help="restore checkpoint", default=r"D:\jx\ai\stereo_match_dl\IGEV-main\IGEV-main\IGEV-Stereo\ckpt\41481_epoch_igev-stereo.pth") # parser.add_argument('-l', '--left_imgs', help="path to all first (left) frames", default=r"D:\jx\ai\stereo_match_dl\test_image\bak\SV1-02_20190416_L1B0000976744_8041900187280009_01-PAN2_epi.tif") # parser.add_argument('-r', '--right_imgs', help="path to all second (right) frames", default=r"D:\jx\ai\stereo_match_dl\test_image\bak\SV1-02_20190416_L1B0000976744_8041900187280009_01-PAN1_epi.tif") parser.add_argument('-l', '--left_imgs', help="path to all first (left) frames", default=r"D:\jx\ai\stereo_match_dl\test_image\bak\SV1_PAN1_down_half_resized.tif") parser.add_argument('-r', '--right_imgs', help="path to all second (right) frames", default=r"D:\jx\ai\stereo_match_dl\test_image\bak\SV1_PAN2_down_half_resized.tif") # parser.add_argument('-l', '--left_imgs', help="path to all first (left) frames", default=r"D:\jx\ai\stereo_match_dl\test_image\right\3_6_right.tif") # parser.add_argument('-r', '--right_imgs', help="path to all second (right) frames", default=r"D:\jx\ai\stereo_match_dl\test_image\left\3_6_left.tif") #parser.add_argument('-l', '--left_imgs', help="path to all first (left) frames", default=r"D:\jx\ai\stereo_match_dl\train_image\experimental data\with ground truth\test\left\HY_left_0.tiff") #parser.add_argument('-r', '--right_imgs', help="path to all second (right) frames", default=r"D:\jx\ai\stereo_match_dl\train_image\experimental data\with ground truth\test\right\HY_right_0.tiff") parser.add_argument('--output_directory', help="directory to save output", default=r"D:\jx\ai\stereo_match_dl\IGEV-main\IGEV-main\IGEV-Stereo\core\demo-output\test_patch") parser.add_argument('--mixed_precision', action='store_true', help='use mixed precision') # Architecture choices parser.add_argument('--hidden_dims', nargs='+', type=int, default=[128]*3, help="hidden state and context dimensions") parser.add_argument('--corr_implementation', choices=["reg", "alt", "reg_cuda", "alt_cuda"], default="reg", help="correlation volume implementation") parser.add_argument('--shared_backbone', action='store_true', help="use a single backbone for the context and feature encoders") parser.add_argument('--corr_levels', type=int, default=2, help="number of levels in the correlation pyramid") parser.add_argument('--corr_radius', type=int, default=4, help="width of the correlation pyramid") parser.add_argument('--n_downsample', type=int, default=2, help="resolution of the disparity field (1/2^K)") parser.add_argument('--slow_fast_gru', action='store_true', help="iterate the low-res GRUs more frequently") parser.add_argument('--n_gru_layers', type=int, default=3, help="number of hidden GRU levels") parser.add_argument('--max_disp', type=int, default=192, help="max disp of geometry encoding volume") args = parser.parse_args() Path(args.output_directory).mkdir(exist_ok=True, parents=True) # 初始化模型 model = torch.nn.DataParallel(IGEVStereo(args), device_ids=[0]) model.load_state_dict(torch.load(args.restore_ckpt)) model = model.module model.to(DEVICE) model.eval() result = get_init_disp(args.left_imgs, args.right_imgs, model) get_disp_img_pyramid(model, args.left_imgs, args.right_imgs, result, args.output_directory)关键修改点说明:
-
创建权重矩阵函数:
def create_weight_mask(height, width): """ 创建一个线性权重矩阵,中心权重最大,边缘权重最小。 这里使用简单的线性渐变,也可以使用高斯权重。 """ # 创建水平和垂直方向的线性权重 weight_x = np.linspace(0, 1, width) weight_y = np.linspace(0, 1, height) weight = np.outer(weight_y, weight_x) # 归一化权重到 [0,1] weight = weight / np.max(weight) return weight.astype(np.float32)该函数生成一个线性权重矩阵,可根据需要调整为高斯权重或其他形式的权重矩阵。
-
应用权重矩阵: 在处理每个块时,生成与
disp_patch_non_expanded大小匹配的权重矩阵,并将其应用于视差值的累加:# 创建与 disp_patch_non_expanded 尺寸匹配的权重矩阵 mask = create_weight_mask(disp_patch_non_expanded.shape[0], disp_patch_non_expanded.shape[1]) # 将权重应用于 disp_patch_non_expanded disp_map[l_h_start:l_h_end, l_w_start:l_w_end] += disp_patch_non_expanded * mask weight_map[l_h_start:l_h_end, l_w_start:l_w_end] += mask + skew_disp -
最终归一化: 在所有块处理完成后,对累积的视差图进行归一化:
# 归一化 disparity map non_zero_mask = weight_map > 0 disp_map[non_zero_mask] = disp_map[non_zero_mask] / weight_map[non_zero_mask] # 将无效区域设为0或其他默认值 disp_map[~non_zero_mask] = 0 -
保存视差图: 为确保保存的视差图数据类型正确,建议使用
.astype(np.float32):imageio.imwrite(filename, disp_map.astype(np.float32))
其他建议:
-
权重矩阵的选择: 你可以尝试不同类型的权重矩阵,如高斯权重,以获得更平滑的过渡效果。例如,使用高斯滤波权重:
from scipy.ndimage import gaussian_filter def create_gaussian_weight_mask(height, width, sigma=width//4): """ 创建一个高斯权重矩阵,中心权重最高,边缘权重逐渐减小。 """ x = np.linspace(-1, 1, width) y = np.linspace(-1, 1, height) xv, yv = np.meshgrid(x, y) weight = np.exp(-(xv**2 + yv**2) / (2 * (sigma / width)**2)) weight = weight / np.max(weight) return weight.astype(np.float32) -
边缘处理: 在边缘区域,由于没有重叠的块,可以考虑使用不完全的权重矩阵,确保不会引入边缘伪影。
-
性能优化: 根据数据量和计算资源,可以调整块大小和权重矩阵生成方式,以平衡性能和效果。
通过上述修改,引入权重矩阵可以有效减少块之间的拼缝现象,获得更平滑和连续的视差图。根据具体的应用场景和图像特性,您可以进一步调整权重矩阵的类型和参数,以达到最佳效果。
-