app.py

import os
import cv2
import torch
import numpy as np
import gradio as gr

import trimesh
import sys
import os

sys.path.append('vggsfm_code/')
import shutil
from datetime import datetime

from vggsfm_code.hf_demo import demo_fn
from omegaconf import DictConfig, OmegaConf
from viz_utils.viz_fn import add_camera, apply_density_filter_np
import glob
# 
from scipy.spatial.transform import Rotation
# import PIL
import gc
import open3d as o3d
import time
# import spaces

# @spaces.GPU
def vggsfm_demo(
    input_video,
    input_image,
    query_frame_num,
    max_query_pts=4096,
):

    
    start_time = time.time()
    gc.collect()
    torch.cuda.empty_cache()

    debug = False

    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")

    max_input_image = 25

    target_dir = f"input_images_{timestamp}"
    if os.path.exists(target_dir): 
        shutil.rmtree(target_dir)

    os.makedirs(target_dir)
    target_dir_images = target_dir + "/images"
    os.makedirs(target_dir_images)


    if debug:
        predictions = torch.load("predictions_scene2.pth")
    else:
        
        if input_video is not None:            
            if not isinstance(input_video, str):
                input_video = input_video["video"]["path"]
        
        cfg_file = "vggsfm_code/cfgs/demo.yaml"
        cfg = OmegaConf.load(cfg_file)

        if input_image is not None:

            input_image = sorted(input_image)
            input_image = input_image[:max_input_image]
            
            recon_num = len(input_image)
            if recon_num<3:
                return None, "Please input at least three frames"

            # Copy files to the new directory
            for file_name in input_image:
                shutil.copy(file_name, target_dir_images)
        elif input_video is not None:
            vs = cv2.VideoCapture(input_video)

            fps = vs.get(cv2.CAP_PROP_FPS)

            frame_rate = 1
            frame_interval = int(fps * frame_rate)
            
            video_frame_num = 0
            count = 0 
            
            while video_frame_num<max_input_image:
                (gotit, frame) = vs.read()
                count +=1

                if not gotit:
                    break
                
                if count % frame_interval == 0:
                    cv2.imwrite(target_dir_images+"/"+f"{video_frame_num:06}.png", frame)
                    video_frame_num+=1
                    
            recon_num = video_frame_num     
            if recon_num<3:
                return None, "Please input at least three frames"
        else:
            return None, "Uploading not finished or Incorrect input format"
            
        cfg.query_frame_num = query_frame_num
        cfg.max_query_pts = max_query_pts
        print(f"Files have been copied to {target_dir_images}")
        cfg.SCENE_DIR = target_dir
        
        # try:
        predictions = demo_fn(cfg)
        # except:
        # return None, "Something seems to be incorrect. Please verify that your inputs are formatted correctly. If the issue persists, kindly create a GitHub issue for further assistance."
    
    glbscene = vggsfm_predictions_to_glb(predictions)
    
    glbfile = target_dir + "/glbscene.glb"
    glbscene.export(file_obj=glbfile) 
    # glbscene.export(file_obj=glbfile, line_settings= {'point_size': 20})    


    del predictions
    gc.collect()
    torch.cuda.empty_cache()
    
    print(input_image)
    print(input_video)
    end_time = time.time()
    execution_time = end_time - start_time
    print(f"Execution time: {execution_time} seconds")

    # recon_num
    return glbfile, f"Reconstruction complete ({recon_num} frames)"




def vggsfm_predictions_to_glb(predictions, sphere=False):
    # del predictions['reconstruction']
    # torch.save(predictions, "predictions_scene2.pth")
    # learned from https://github.com/naver/dust3r/blob/main/dust3r/viz.py
    points3D = predictions["points3D"].cpu().numpy()
    points3D_rgb = predictions["points3D_rgb"].cpu().numpy()
    points3D_rgb = (points3D_rgb*255).astype(np.uint8)
    
    extrinsics_opencv = predictions["extrinsics_opencv"].cpu().numpy()
    intrinsics_opencv = predictions["intrinsics_opencv"].cpu().numpy()
    
        
    raw_image_paths = predictions["raw_image_paths"]
    images = predictions["images"].permute(0,2,3,1).cpu().numpy()
    images = (images*255).astype(np.uint8)
    
    glbscene = trimesh.Scene()
    
    if True:
        pcd = o3d.geometry.PointCloud()
        pcd.points = o3d.utility.Vector3dVector(points3D)
        pcd.colors = o3d.utility.Vector3dVector(points3D_rgb)
        
        cl, ind = pcd.remove_statistical_outlier(nb_neighbors=20, std_ratio=1.0)
        filtered_pcd = pcd.select_by_index(ind)

        print(f"Filter out {len(points3D) - len(filtered_pcd.points)} 3D points")
        points3D = np.asarray(filtered_pcd.points)
        points3D_rgb = np.asarray(filtered_pcd.colors)

    
    if sphere:
        # TOO SLOW
        print("testing sphere")
        # point_size = 0.02  
    else: 
        point_cloud = trimesh.PointCloud(points3D, colors=points3D_rgb)
        glbscene.add_geometry(point_cloud)


    camera_edge_colors = [(255, 0, 0), (0, 0, 255), (0, 255, 0), (255, 0, 255), (255, 204, 0), (0, 204, 204),
                (128, 255, 255), (255, 128, 255), (255, 255, 128), (0, 0, 0), (128, 128, 128)]

    frame_num = len(extrinsics_opencv)
    extrinsics_opencv_4x4 = np.zeros((frame_num, 4, 4))
    extrinsics_opencv_4x4[:, :3, :4] = extrinsics_opencv
    extrinsics_opencv_4x4[:, 3, 3] = 1

    for idx in range(frame_num):
        cam_from_world = extrinsics_opencv_4x4[idx]
        cam_to_world = np.linalg.inv(cam_from_world)
        cur_cam_color = camera_edge_colors[idx % len(camera_edge_colors)]
        cur_focal = intrinsics_opencv[idx, 0, 0]

        add_camera(glbscene, cam_to_world, cur_cam_color, image=None, imsize=(1024,1024), 
                   focal=None,screen_width=0.35)

    opengl_mat = np.array([[1, 0, 0, 0],
                    [0, -1, 0, 0],
                    [0, 0, -1, 0],
                    [0, 0, 0, 1]])

    rot = np.eye(4)
    rot[:3, :3] = Rotation.from_euler('y', np.deg2rad(180)).as_matrix()
    glbscene.apply_transform(np.linalg.inv(np.linalg.inv(extrinsics_opencv_4x4[0]) @ opengl_mat @ rot))

    # Calculate the bounding box center and apply the translation
    # bounding_box = glbscene.bounds
    # center = (bounding_box[0] + bounding_box[1]) / 2
    # translation = np.eye(4)
    # translation[:3, 3] = -center

    # glbscene.apply_transform(translation)
    # glbfile = "glbscene.glb"
    # glbscene.export(file_obj=glbfile)    
    return glbscene


statue_video = "vggsfm_code/examples/videos/statue_video.mp4"

apple_video = "vggsfm_code/examples/videos/apple_video.mp4"
british_museum_video = "vggsfm_code/examples/videos/british_museum_video.mp4"
cake_video = "vggsfm_code/examples/videos/cake_video.mp4"
bonsai_video = "vggsfm_code/examples/videos/bonsai_video.mp4"
face_video =  "vggsfm_code/examples/videos/in2n_face_video.mp4"
counter_video =  "vggsfm_code/examples/videos/in2n_counter_video.mp4"

horns_video = "vggsfm_code/examples/videos/llff_horns_video.mp4"
person_video = "vggsfm_code/examples/videos/in2n_person_video.mp4"

flower_video = "vggsfm_code/examples/videos/llff_flower_video.mp4"

fern_video = "vggsfm_code/examples/videos/llff_fern_video.mp4"


drums_video = "vggsfm_code/examples/videos/drums_video.mp4"

kitchen_video = "vggsfm_code/examples/videos/kitchen_video.mp4"

###########################################################################################
apple_images = glob.glob(f'vggsfm_code/examples/apple/images/*')
bonsai_images = glob.glob(f'vggsfm_code/examples/bonsai/images/*')
cake_images = glob.glob(f'vggsfm_code/examples/cake/images/*')
british_museum_images = glob.glob(f'vggsfm_code/examples/british_museum/images/*')
face_images = glob.glob(f'vggsfm_code/examples/in2n_face/images/*')
counter_images = glob.glob(f'vggsfm_code/examples/in2n_counter/images/*')

horns_images = glob.glob(f'vggsfm_code/examples/llff_horns/images/*')

person_images = glob.glob(f'vggsfm_code/examples/in2n_person/images/*')
flower_images = glob.glob(f'vggsfm_code/examples/llff_flower/images/*')

fern_images = glob.glob(f'vggsfm_code/examples/llff_fern/images/*')
statue_images = glob.glob(f'vggsfm_code/examples/statue/images/*')

drums_images = glob.glob(f'vggsfm_code/examples/drums/images/*')
kitchen_images = glob.glob(f'vggsfm_code/examples/kitchen/images/*')



###########################################################################################


with gr.Blocks() as demo:
    
    gr.Markdown("""
    # 🏛️ VGGSfM: Visual Geometry Grounded Deep Structure From Motion
    
    <div style="font-size: 16px; line-height: 1.2;">
    Welcome to <a href="https://vggsfm.github.io/" target="_blank" style="color: #2a9d8f;">VGGSfM</a> 🤗 demo! This space demonstrates 3D reconstruction from input image frames.
    
    <h3 style="color: #2a9d8f;">Get Started</h3>
    To get started quickly, you can click on our <strong>examples (at the bottom of the page)</strong>. The example results are cached, allowing you to view them even when in a queue.

    If you want to reconstruct your own data, simply **(a)** upload images (.jpg, .png, etc.) or **(b)** upload a video (.mp4, .mov, etc.).

    <div style="font-size: 16px; line-height: 1.2;">
    <h3 style="color: #2a9d8f;">Hyperparameters</h3>
    Typically, 4 query images and 2048 query points are sufficient. For a denser point cloud, use 4096 query points. If the reconstruction appears incomplete, increase to 6 query images. Note that excessive queries can lead to out-of-memory errors.
    
    <h3 style="color: #2a9d8f;">Converting Video to Frames</h3>
    By default, we convert the input video to frames at 1 frame per second. To prevent hugging face space crashes, we limit reconstruction to the first 25 frames. If both images and videos are uploaded, the demo will only reconstruct the uploaded images.

    <h3 style="color: #2a9d8f;">Dynamic Scenes</h3>
    SfM methods are designed for rigid/static reconstruction. When dealing with dynamic/moving inputs, these methods may still work by focusing on the rigid parts of the scene. However, to ensure high-quality results, it is better to minimize the presence of moving objects in the input data.

    <h3 style="color: #2a9d8f;">Runtime</h3>
    The reconstruction typically takes <strong>up to 90 seconds</strong>. Longer runtimes can be attributed to difficult input data or a high number of query images/points. Please note that running reconstruction on Hugging Face is slower than on a local machine.

    <h3 style="color: #2a9d8f;">Contact</h3>
    If you meet any problem, feel free to create an issue in our <a href="https://github.com/facebookresearch/vggsfm" target="_blank" style="color: #2a9d8f;">GitHub Repo</a> ⭐
    </div>
    """)


    with gr.Row():
        with gr.Column(scale=1):
            input_video = gr.Video(label="Upload Video", interactive=True)
            input_images = gr.File(file_count="multiple", label="Upload Images", interactive=True)
            num_query_images = gr.Slider(minimum=1, maximum=10, step=1, value=4, label="Number of query images (key frames)",
                                         info="More query images usually lead to better reconstruction at a lower speed. If the viewpoint differences between your images are minimal, you can set this value to 1. ")
            num_query_points = gr.Slider(minimum=600, maximum=6000, step=1, value=2048, label="Number of query points",
                                         info="More query points usually lead to denser reconstruction at a lower speed.")
        
        with gr.Column(scale=3):
            reconstruction_output = gr.Model3D(label="3D Reconstruction (Point Cloud and Camera Poses; Zoom in to see details)", height=520, zoom_speed=0.5, pan_speed=0.5)
            log_output = gr.Textbox(label="Log")

    with gr.Row():
        submit_btn = gr.Button("Reconstruct", scale=1)

        # submit_btn = gr.Button("Reconstruct", scale=1, elem_attributes={"style": "background-color: blue; color: white;"})
        clear_btn = gr.ClearButton([input_video, input_images, num_query_images, num_query_points, reconstruction_output, log_output], scale=1)
    
    
    
    
    examples = [
        [flower_video, flower_images, 2, 4096],
        [kitchen_video, kitchen_images, 4, 2048],
        [person_video, person_images, 3, 2048],
        [statue_video, statue_images, 4, 2048],
        [drums_video, drums_images, 4, 2048],
        [counter_video, counter_images, 4, 2048],
        [fern_video, fern_images, 2, 4096],
        [horns_video, horns_images, 3, 4096],
        [apple_video, apple_images, 6, 2048],
        [british_museum_video, british_museum_images, 1, 4096],
        [bonsai_video, bonsai_images, 3, 2048],
        # [face_video, face_images, 4, 2048],
        # [cake_video, cake_images, 3, 2048],
    ]
    
    
    
    gr.Examples(examples=examples, 
                inputs=[input_video, input_images, num_query_images, num_query_points],
                outputs=[reconstruction_output, log_output],  # Provide outputs
                fn=vggsfm_demo,  # Provide the function
                cache_examples=True,
                examples_per_page=30,
                )


    submit_btn.click(
        vggsfm_demo,
        [input_video, input_images, num_query_images, num_query_points],
        [reconstruction_output, log_output],
        concurrency_limit=1
    )

    # demo.launch(debug=True, share=True)
    demo.queue(max_size=20).launch(show_error=True, share=True)
    # demo.queue(max_size=20, concurrency_count=1).launch(debug=True, share=True)
########################################################################################################################