import sys sys.path.append("flash3d") from omegaconf import OmegaConf import gradio as gr import spaces import torch import torchvision.transforms as TT import torchvision.transforms.functional as TTF from huggingface_hub import hf_hub_download from networks.gaussian_predictor import GaussianPredictor from util.vis3d import save_ply def main(): if torch.cuda.is_available(): device = "cuda:0" else: device = "cpu" model_cfg_path = hf_hub_download(repo_id="einsafutdinov/flash3d", filename="config_re10k_v1.yaml") model_path = hf_hub_download(repo_id="einsafutdinov/flash3d", filename="model_re10k_v1.pth") cfg = OmegaConf.load(model_cfg_path) model = GaussianPredictor(cfg) device = torch.device(device) model.load_model(model_path) model.to(device) pad_border_fn = TT.Pad((cfg.dataset.pad_border_aug, cfg.dataset.pad_border_aug)) to_tensor = TT.ToTensor() def check_input_image(input_image): if input_image is None: raise gr.Error("No image uploaded!") def preprocess(image, dynamic_size=False, padding=True): h, w = image.size size = 32 if dynamic_size: while max(h, w) // size > 20: size *=2 crop_image = TTF.center_crop(image, (w // size * size, h // size * size)) resize_image = TTF.resize(crop_image, (w // size * 32, h // size * 32), interpolation=TT.InterpolationMode.BICUBIC) model.cfg.dataset.width, model.cfg.dataset.height = resize_image.size else: model.cfg.dataset.height, model.cfg.dataset.width = 256, 384 resize_image = TTF.resize( image, (model.cfg.dataset.height, model.cfg.dataset.width), interpolation=TT.InterpolationMode.BICUBIC ) if padding: input_image = pad_border_fn(resize_image) model.cfg.dataset.pad_border_aug = 32 else: input_image = resize_image model.cfg.dataset.pad_border_aug = 0 model.set_backproject() return input_image @spaces.GPU() def reconstruct_and_export(image): """ Passes image through model, outputs reconstruction in form of a dict of tensors. """ image = to_tensor(image).to(device).unsqueeze(0) inputs = { ("color_aug", 0, 0): image, } outputs = model(inputs) # export reconstruction to ply save_ply(outputs, ply_out_path, num_gauss=model.cfg.model.gaussians_per_pixel, h=model.cfg.dataset.height, w=model.cfg.dataset.width, pad=model.cfg.dataset.pad_border_aug) return ply_out_path ply_out_path = f'./mesh.ply' css = """ h1 { text-align: center; display:block; } """ with gr.Blocks(css=css) as demo: gr.Markdown( """ # Flash3D **Flash3D** [[project page](https://www.robots.ox.ac.uk/~vgg/research/flash3d/)] is a fast, super efficient, trainable on a single GPU in a day for scene 3D reconstruction from a single image. The model used in the demo was trained on only **RealEstate10k dataset on a single A6000 GPU within 1 day**. Upload an image of a scene or click on one of the provided examples to see how the Flash3D does. The 3D viewer will render a .ply scene exported from the 3D Gaussians, which is only an approximation. """ ) with gr.Row(variant="panel"): with gr.Column(scale=1): with gr.Row(): input_image = gr.Image( label="Input Image", image_mode="RGBA", sources="upload", type="pil", elem_id="content_image", ) with gr.Row(): submit = gr.Button("Generate", elem_id="generate", variant="primary") with gr.Row(): dynamic_size = gr.Checkbox(True, interactive=True, label='Use the original image ratio') padding = gr.Checkbox(True, interactive=True, label='add padding to the image') with gr.Row(variant="panel"): gr.Examples( examples=[ './demo_examples/bedroom_01.png', './demo_examples/kitti_02.png', './demo_examples/kitti_03.png', './demo_examples/re10k_05.jpg', './demo_examples/re10k_06.jpg', './demo_examples/christ_church_cathedral.png', './demo_examples/radcliffe.png', './demo_examples/blenheim_palace_bedroom.png', './demo_examples/blenheim_palace_living.png', './demo_examples/blenheim_palace.JPG', ], inputs=[input_image], cache_examples=False, label="Examples", examples_per_page=20, ) with gr.Row(): processed_image = gr.Image(label="Processed Image", interactive=False) with gr.Column(scale=2): with gr.Row(): with gr.Tab("Reconstruction"): output_model = gr.Model3D( height=640, label="Output Model", interactive=False ) gr.Markdown( """ ## Comments: 1. If you run the demo online, the first example you upload should take about 25 seconds (with preprocessing, saving and overhead), the following take about 14s (due to the .ply visualisation). 2. The 3D viewer shows a .ply mesh extracted from a mix of 3D Gaussians. This is only an approximations and artefacts might show. 3. Known limitations include: - a black dot appearing on the model from some viewpoints - while the multiple gaussians fill in resonable pixels to the invisible parts, the visual quality is still blurry. 4. It achieves state-of-the-art results when trained and tested on RealEstate10k., and is **much** cheaper to train and run. 5. When transferred to unseen datasets like NYU it outperforms competitors by a large margin. 6. More impressively, when transferred to KITTI, Flash3D achieves better PSNR than methods trained specifically on that dataset. ## How does it work? Given a single image I as input, Flash3D first estimates the metric depth D using a frozen off-the-shelf network. Then, a ResNet50-like encoder–decoder network predicts a set of shape and appearance parameters P of K layers of Gaussians for every pixel u, allowing unobserved and occluded surfaces to be modelled. From these predicted components, the depth can be obtained by summing the predicted (positive) offsets δi with the predicted monocular depth D, allowing the mean vector for every layer of Gaussians to be computed. This strategy ensures that the layers are depth-ordered, encouraging the network to model occluded surfaces. For more results see the [project page](https://www.robots.ox.ac.uk/~vgg/research/flash3d/). """ ) submit.click(fn=check_input_image, inputs=[input_image]).success( fn=preprocess, inputs=[input_image, dynamic_size, padding], outputs=[processed_image], ).success( fn=reconstruct_and_export, inputs=[processed_image], outputs=[output_model], ) demo.queue(max_size=1) demo.launch() if __name__ == "__main__": main()