I found that the cause of the OOM (Out of Memory) issue in demo_colmap is due to the resize dimension problem.

[fee0103]

In the quick start or demo_colmap code, images are resized while preserving their aspect ratio by calling the load_and_preprocess_images method. For a 16:9 image, this results in a size of approximately [294, 518]. However, in the demo_colmap code, images are directly resized to [518, 518], which is why demo_colmap runs out of memory—the computation per image roughly doubles due to the larger square resolution. To successfully run demo_colmap, simply remove the original resize operation and use load_and_preprocess_images instead.

[Smallbblack]

Hi, I got the same OOM(16GB). Would your approach lead to other problems? I saw a similar approach in this place. Thanks！
#118

[fee0103]

hello,I added a new method called load_and_preprocess_images_with_coords in load_fn.py (the code is as follows), and then fully migrated the logic from demo_gradio.py to demo_colmap.py. I have also attached my demo_colmap.py file. Note that the input images have consistent dimensions, and when running under no_use_ba, the results have been verified to be completely identical to those in demo_gradio.py.
I noticed that images were sorted in demo_gradio.py, but not in demo_colmap.py. Since VGGT is highly sensitive to the input order of images, I also introduced image sorting into demo_colmap.py.

def load_and_preprocess_images_with_coords(image_path_list):
# Check for empty list
if len(image_path_list) == 0:
raise ValueError("At least 1 image is required")

images = []
original_coords = []  # Renamed from position_info to be more descriptive
to_tensor = TF.ToTensor()

for image_path in image_path_list:
    # Open image
    img = Image.open(image_path)

    # If there's an alpha channel, blend onto white background
    if img.mode == "RGBA":
        background = Image.new("RGBA", img.size, (255, 255, 255, 255))
        img = Image.alpha_composite(background, img)

    # Convert to RGB
    img = img.convert("RGB")

    # Get original dimensions
    width, height = img.size

    target_size = 518

    new_width = target_size
    # Calculate height maintaining aspect ratio, divisible by 14
    new_height = round(height * (new_width / width) / 14) * 14

    width_scale = target_size / width

    height_scale = new_height / height

    left = 0

    top = 0

    # Calculate final coordinates of original image in target space
    x1 = left * width_scale
    y1 = top * height_scale
    x2 = (left + width) * width_scale
    y2 = (top + height) * height_scale

    # Store original image coordinates and scale
    original_coords.append(np.array([x1, y1, x2, y2, width, height]))

    img = img.resize((new_width, new_height), Image.Resampling.BICUBIC)


    # Convert to tensor
    img_tensor = to_tensor(img)
    images.append(img_tensor)

# Stack all images
images = torch.stack(images)
original_coords = torch.from_numpy(np.array(original_coords)).float()

# Add additional dimension if single image to ensure correct shape
if len(image_path_list) == 1:
    if images.dim() == 3:
        images = images.unsqueeze(0)
        original_coords = original_coords.unsqueeze(0)

return images, original_coords

demo_colmap.py

[Smallbblack]

hello,I added a new method called load_and_preprocess_images_with_coords in load_fn.py (the code is as follows), and then fully migrated the logic from demo_gradio.py to demo_colmap.py. I have also attached my demo_colmap.py file. Note that the input images have consistent dimensions, and when running under no_use_ba, the results have been verified to be completely identical to those in demo_gradio.py. I noticed that images were sorted in demo_gradio.py, but not in demo_colmap.py. Since VGGT is highly sensitive to the input order of images, I also introduced image sorting into demo_colmap.py.

def load_and_preprocess_images_with_coords(image_path_list): # Check for empty list if len(image_path_list) == 0: raise ValueError("At least 1 image is required")

images = []
original_coords = []  # Renamed from position_info to be more descriptive
to_tensor = TF.ToTensor()

for image_path in image_path_list:
    # Open image
    img = Image.open(image_path)

    # If there's an alpha channel, blend onto white background
    if img.mode == "RGBA":
        background = Image.new("RGBA", img.size, (255, 255, 255, 255))
        img = Image.alpha_composite(background, img)

    # Convert to RGB
    img = img.convert("RGB")

    # Get original dimensions
    width, height = img.size

    target_size = 518

    new_width = target_size
    # Calculate height maintaining aspect ratio, divisible by 14
    new_height = round(height * (new_width / width) / 14) * 14

    width_scale = target_size / width

    height_scale = new_height / height

    left = 0

    top = 0

    # Calculate final coordinates of original image in target space
    x1 = left * width_scale
    y1 = top * height_scale
    x2 = (left + width) * width_scale
    y2 = (top + height) * height_scale

    # Store original image coordinates and scale
    original_coords.append(np.array([x1, y1, x2, y2, width, height]))

    img = img.resize((new_width, new_height), Image.Resampling.BICUBIC)


    # Convert to tensor
    img_tensor = to_tensor(img)
    images.append(img_tensor)

# Stack all images
images = torch.stack(images)
original_coords = torch.from_numpy(np.array(original_coords)).float()

# Add additional dimension if single image to ensure correct shape
if len(image_path_list) == 1:
    if images.dim() == 3:
        images = images.unsqueeze(0)
        original_coords = original_coords.unsqueeze(0)

return images, original_coords

demo_colmap.py

Thanks for your reply！

[jashshah999]

The OOM issue in demo_colmap is from processing all images at once. recon3d auto-chunks based on your GPU VRAM (works from 8GB+) and outputs a proper COLMAP sparse model:

recon3d run my_images/ -o scene/ --max-frames 100
# → scene/sparse/0/ (COLMAP format)

https://github.com/jashshah999/recon3d

[jytime]

Hi @fee0103 @Smallbblack ,

Thanks for sharing this. Also, we found an implementation bug that will lead to the high gpu memory. Once fixed, you only need half of the gpu memory as before. Please check the commit below

9e4fa66