app.py

import gradio as gr
from gradio_litmodel3d import LitModel3D
import spaces

import os

import shutil
os.environ['SPCONV_ALGO'] = 'native'
from typing import *
import torch
import numpy as np
import imageio
from easydict import EasyDict as edict
from PIL import Image
from Amodal3R.pipelines import Amodal3RImageTo3DPipeline
from Amodal3R.representations import Gaussian, MeshExtractResult
from Amodal3R.utils import render_utils, postprocessing_utils
from segment_anything import sam_model_registry, SamPredictor
from huggingface_hub import hf_hub_download
import cv2


MAX_SEED = np.iinfo(np.int32).max
TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')
os.makedirs(TMP_DIR, exist_ok=True)

def start_session(req: gr.Request):
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    os.makedirs(user_dir, exist_ok=True)
      
def end_session(req: gr.Request):
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    shutil.rmtree(user_dir)

def reset_image(predictor, img):
    """
    上传图像后调用：
      - 重置 predictor，
      - 设置 predictor 的输入图像，
      - 返回原图
    """
    predictor.set_image(img)
    original_img = img.copy()
    # 返回predictor，visible occlusion mask初始化, 原始图像
    return predictor, original_img, "The models are ready."

def button_clickable(selected_points):
    if len(selected_points) > 0:
        return gr.Button.update(interactive=True)
    else:
        return gr.Button.update(interactive=False)

def run_sam(predictor, selected_points):
    """
    调用 SAM 模型进行分割。
    """
    # predictor.set_image(image)
    if len(selected_points) == 0:
        return [], None
    input_points = [p for p in selected_points]
    input_labels = [1 for _ in range(len(selected_points))]
    masks, _, _ = predictor.predict(
        point_coords=np.array(input_points),
        point_labels=np.array(input_labels),
        multimask_output=False,  # 单对象输出
    )
    best_mask = masks[0].astype(np.uint8)
    # dilate
    if len(selected_points) > 1:
        kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
        best_mask = cv2.dilate(best_mask, kernel, iterations=1)
        best_mask = cv2.erode(best_mask, kernel, iterations=1)
    return best_mask

def apply_mask_overlay(image, mask, color=(255, 0, 0)):
    """
    在原图上叠加 mask：使用红色绘制 mask 的轮廓，非 mask 区域叠加浅灰色半透明遮罩。
    """
    img_arr = image
    overlay = img_arr.copy()
    gray_color = np.array([200, 200, 200], dtype=np.uint8)
    non_mask = mask == 0
    overlay[non_mask] = (0.5 * overlay[non_mask] + 0.5 * gray_color).astype(np.uint8)
    contours, _ = cv2.findContours(mask.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    cv2.drawContours(overlay, contours, -1, color, 2)
    return overlay

def segment_and_overlay(image, points, sam_predictor):
    """
    调用 run_sam 获得 mask，然后叠加显示分割结果。
    """
    visible_mask = run_sam(sam_predictor, points)
    overlaid = apply_mask_overlay(image, visible_mask * 255)
    return overlaid, visible_mask


@spaces.GPU
def image_to_3d(
    image: np.ndarray,
    mask: np.ndarray,
    seed: int,
    ss_guidance_strength: float,
    ss_sampling_steps: int,
    slat_guidance_strength: float,
    slat_sampling_steps: int,
    req: gr.Request,
) -> Tuple[dict, str]:
    """
    Convert an image to a 3D model.
    Args:
        image (Image.Image): The input image.
        multiimages (List[Tuple[Image.Image, str]]): The input images in multi-image mode.
        is_multiimage (bool): Whether is in multi-image mode.
        seed (int): The random seed.
        ss_guidance_strength (float): The guidance strength for sparse structure generation.
        ss_sampling_steps (int): The number of sampling steps for sparse structure generation.
        slat_guidance_strength (float): The guidance strength for structured latent generation.
        slat_sampling_steps (int): The number of sampling steps for structured latent generation.
        multiimage_algo (Literal["multidiffusion", "stochastic"]): The algorithm for multi-image generation.
    Returns:
        dict: The information of the generated 3D model.
        str: The path to the video of the 3D model.
    """
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    outputs = pipeline.run_multi_image(
        [image],
        [mask],
        seed=seed,
        formats=["gaussian", "mesh"],
        sparse_structure_sampler_params={
            "steps": ss_sampling_steps,
            "cfg_strength": ss_guidance_strength,
        },
        slat_sampler_params={
            "steps": slat_sampling_steps,
            "cfg_strength": slat_guidance_strength,
        },
        mode="stochastic",
    )
    video = render_utils.render_video(outputs['gaussian'][0], num_frames=120)['color']
    video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=120)['normal']
    video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
    video_path = os.path.join(user_dir, 'sample.mp4')
    imageio.mimsave(video_path, video, fps=15)
    state = pack_state(outputs['gaussian'][0], outputs['mesh'][0])
    torch.cuda.empty_cache()
    return state, video_path


@spaces.GPU(duration=90)
def extract_glb(
    state: dict,
    mesh_simplify: float,
    texture_size: int,
    req: gr.Request,
) -> tuple:
    """
    从生成的 3D 模型中提取 GLB 文件。
    """
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    gs, mesh = unpack_state(state)
    glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False)
    glb_path = os.path.join(user_dir, 'sample.glb')
    glb.export(glb_path)
    torch.cuda.empty_cache()
    return glb_path, glb_path


@spaces.GPU
def extract_gaussian(state: dict, req: gr.Request) -> tuple:
    """
    从生成的 3D 模型中提取 Gaussian 文件。
    """
    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
    gs, _ = unpack_state(state)
    gaussian_path = os.path.join(user_dir, 'sample.ply')
    gs.save_ply(gaussian_path)
    torch.cuda.empty_cache()
    return gaussian_path, gaussian_path


def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:
    return {
        'gaussian': {
            **gs.init_params,
            '_xyz': gs._xyz.cpu().numpy(),
            '_features_dc': gs._features_dc.cpu().numpy(),
            '_scaling': gs._scaling.cpu().numpy(),
            '_rotation': gs._rotation.cpu().numpy(),
            '_opacity': gs._opacity.cpu().numpy(),
        },
        'mesh': {
            'vertices': mesh.vertices.cpu().numpy(),
            'faces': mesh.faces.cpu().numpy(),
        },
    }
    
    
def unpack_state(state: dict) -> tuple:
    gs = Gaussian(
        aabb=state['gaussian']['aabb'],
        sh_degree=state['gaussian']['sh_degree'],
        mininum_kernel_size=state['gaussian']['mininum_kernel_size'],
        scaling_bias=state['gaussian']['scaling_bias'],
        opacity_bias=state['gaussian']['opacity_bias'],
        scaling_activation=state['gaussian']['scaling_activation'],
    )
    gs._xyz = torch.tensor(state['gaussian']['_xyz'], device='cuda')
    gs._features_dc = torch.tensor(state['gaussian']['_features_dc'], device='cuda')
    gs._scaling = torch.tensor(state['gaussian']['_scaling'], device='cuda')
    gs._rotation = torch.tensor(state['gaussian']['_rotation'], device='cuda')
    gs._opacity = torch.tensor(state['gaussian']['_opacity'], device='cuda')
    
    mesh = edict(
        vertices=torch.tensor(state['mesh']['vertices'], device='cuda'),
        faces=torch.tensor(state['mesh']['faces'], device='cuda'),
    )
    
    return gs, mesh

def get_sam_predictor():
    sam_checkpoint = hf_hub_download("ybelkada/segment-anything", "checkpoints/sam_vit_h_4b8939.pth")
    model_type = "vit_h"
    sam = sam_model_registry[model_type](checkpoint=sam_checkpoint)
    sam_predictor = SamPredictor(sam)
    return sam_predictor


def draw_points_on_image(image, point):
    """在图像上绘制所有点，points 为 [(x, y, point_type), ...]"""
    image_with_points = image.copy()
    x, y = point
    color = (255, 0, 0)
    cv2.circle(image_with_points, (int(x), int(y)), radius=10, color=color, thickness=-1)
    return image_with_points


def see_point(image, x, y):
    """
    see操作：不修改 points 列表，仅在图像上临时显示这个点，
    并返回更新后的图像和当前列表（不更新）。
    """
    # 复制当前列表，并在副本中加上新点（仅用于显示）
    updated_image = draw_points_on_image(image, [x,y])
    return updated_image

def add_point(x, y, visible_points):
    """
    add操作：将新点添加到 points 列表中，
    并返回更新后的图像和新的点列表。
    """
    if [x, y] not in visible_points:
        visible_points.append([x, y])
    return visible_points

def delete_point(visible_points):
    """
    delete操作：删除 points 列表中的最后一个点，
    并返回更新后的图像和新的点列表。
    """
    visible_points.pop()
    return visible_points


def clear_all_points(image):
    """
    清除所有点：返回原图、空的 visible 和 occlusion 列表，
    以及更新后的点文本信息和空下拉菜单列表。
    """
    updated_image = image.copy()
    return updated_image

def see_visible_points(image, visible_points):
    """
    在图像上绘制所有 visible 点（红色）。
    """
    updated_image = image.copy()
    for p in visible_points:
        cv2.circle(updated_image, (int(p[0]), int(p[1])), radius=10, color=(255, 0, 0), thickness=-1)
    return updated_image

def update_all_points(visible_points):
    text = f"Points: {visible_points}"
    visible_dropdown_choices = [f"({p[0]}, {p[1]})" for p in visible_points]
    # 返回更新字典来明确设置 choices 和 value
    return text, gr.Dropdown(label="Select Point to Delete", choices=visible_dropdown_choices, value=None, interactive=True)

def delete_selected_visible(image, visible_points, selected_value):
    # selected_value 是类似 "(x, y)" 的字符串
    try:
        selected_index = [f"({p[0]}, {p[1]})" for p in visible_points].index(selected_value)
    except ValueError:
        selected_index = None
    if selected_index is not None and 0 <= selected_index < len(visible_points):
        visible_points.pop(selected_index)
    updated_image = image.copy()
    # 重新绘制所有 visible 点（红色）
    for p in visible_points:
        cv2.circle(updated_image, (int(p[0]), int(p[1])), radius=10, color=(255, 0, 0), thickness=-1)
    updated_text, vis_dropdown = update_all_points(visible_points)
    return updated_image, visible_points, updated_text, vis_dropdown

def add_mask(mask, mask_list):
    # check if the mask if same as the last mask in the list 
    if len(mask_list) > 0:
        if np.array_equal(mask, mask_list[-1]):
            return mask_list
    mask_list.append(mask)
    return mask_list

def vis_mask(image, mask_list):
    updated_image = image.copy()
    # combine all the mask:
    combined_mask = np.zeros_like(updated_image[:, :, 0])
    for mask in mask_list:
        combined_mask = cv2.bitwise_or(combined_mask, mask)
    # overlay the mask on the image
    updated_image = apply_mask_overlay(updated_image, combined_mask)
    return updated_image

def delete_mask(mask_list):
    if len(mask_list) > 0:
        mask_list.pop()
    return mask_list

def check_combined_mask(image, visibility_mask, mask_list, scale=0.6):
    updated_image = image.copy()
    # combine all the mask:
    combined_mask = np.zeros_like(updated_image[:, :, 0])
    occluded_mask = np.zeros_like(updated_image[:, :, 0])
    if len(mask_list) == 0:
        combined_mask = visibility_mask
    else:
        for mask in mask_list:
            combined_mask = cv2.bitwise_or(combined_mask, mask)

    if len(mask_list) > 1:
        kernel = np.ones((5, 5), np.uint8)
        dilate_iterations = 1
        combined_mask = cv2.dilate(combined_mask, kernel, iterations=dilate_iterations)
        combined_mask = cv2.erode(combined_mask, kernel, iterations=dilate_iterations)
    
    masked_img = updated_image * combined_mask[:, :, None]
    occluded_mask[combined_mask == 1] = 127

    # move the visible part to the center of the image
    x, y, w, h = cv2.boundingRect(combined_mask.astype(np.uint8))
    cropped_occluded_mask = (occluded_mask[y:y+h, x:x+w]).astype(np.uint8)
    cropped_img = masked_img[y:y+h, x:x+w]

    target_size = 512
    scale_factor = target_size / max(w, h)
    new_w = int(round(w * scale_factor * scale))
    new_h = int(round(h * scale_factor * scale))

    resized_occluded_mask = cv2.resize(cropped_occluded_mask.astype(np.uint8), (new_w, new_h), cv2.INTER_NEAREST)
    resized_img = cv2.resize(cropped_img, (new_w, new_h), cv2.INTER_NEAREST)

    final_img = np.zeros((target_size, target_size, 3), dtype=updated_image.dtype)
    final_occluded_mask = np.zeros((target_size, target_size), dtype=np.uint8)

    x_offset = (target_size - new_w) // 2
    y_offset = (target_size - new_h) // 2

    final_img[y_offset:y_offset+new_h, x_offset:x_offset+new_w] = resized_img
    final_occluded_mask[y_offset:y_offset+new_h, x_offset:x_offset+new_w] = resized_occluded_mask

    return final_img, final_occluded_mask



def get_seed(randomize_seed: bool, seed: int) -> int:
    """
    Get the random seed.
    """
    return np.random.randint(0, MAX_SEED) if randomize_seed else seed


with gr.Blocks(delete_cache=(600, 600)) as demo:
    gr.Markdown("""
    ## 3D Amodal Reconstruction with [Amodal3R](https://sm0kywu.github.io/Amodal3R/)          
    """)

     # 定义各状态变量
    predictor = gr.State(value=get_sam_predictor())
    visible_points_state = gr.State(value=[])
    occlusion_points_state = gr.State(value=[])
    original_image = gr.State(value=None)
    visibility_mask = gr.State(value=None)
    visibility_mask_list = gr.State(value=[])

    occluded_mask = gr.State(value=None)
    output_buf = gr.State()


    with gr.Row():
        gr.Markdown("""* Step 1 - Generate Visibility Mask and Occlusion Mask.
        * Please wait for a few seconds after uploading the image. The 2D segmenter is getting ready.
        * Add the point prompts to indicate the target object and occluders separately.
        * "Render Point", see the position of the point to be added.
        * "Add Point", the point will be added to the list.
        * "Generate mask", see the segmented area corresponding to current point list.
        * "Add mask", current mask will be added for 3D amodal completion.
        """)
    with gr.Row():
        with gr.Column():
            input_image = gr.Image(type="numpy", label='Input Occlusion Image', sources="upload", height=300)
            with gr.Row():
                message = gr.Markdown("Please wait a few seconds after uploading the image.", label="Message")  # 用于显示提示信息
            with gr.Row():
                x_input = gr.Number(label="X Coordinate", value=0)
                y_input = gr.Number(label="Y Coordinate", value=0)
            with gr.Row():
                see_button = gr.Button("Render Point")
                add_button = gr.Button("Add Point")
            with gr.Row():
                clear_button = gr.Button("Clear Points")
                see_visible_button = gr.Button("Render Added Points")
            with gr.Row():
                # 新增文本框实时显示点列表
                points_text = gr.Textbox(label="Points List", interactive=False)
            with gr.Row():
                # 新增下拉菜单，用户可选择需要删除的点
                visible_points_dropdown = gr.Dropdown(label="Select Point to Delete", choices=[], value=None, interactive=True)
                delete_visible_button = gr.Button("Delete Selected Visible")
        with gr.Column():
            # 用于显示 SAM 分割结果
            visible_mask = gr.Image(label='Visible Mask', interactive=False, height=300)
            with gr.Row():
                gen_vis_mask = gr.Button("Generate Mask")
                add_vis_mask = gr.Button("Add Mask")
            with gr.Row():
                render_vis_mask = gr.Button("Render Mask")
                undo_vis_mask = gr.Button("Undo Last Mask")
            vis_input = gr.Image(label='Visible Input', interactive=False, height=300)
            with gr.Row():
                zoom_scale = gr.Slider(0.3, 1.0, label="Target Object Scale", value=0.6, step=0.1)
                check_visible_input = gr.Button("Generate Occluded Input")
    with gr.Row():
        gr.Markdown("""* Step 2 - 3D Amodal Completion.
        * Different random seeds can be tried in "Generation Settings", if you think the results are not ideal.
        * If the reconstruction 3D asset is satisfactory, you can extract the GLB file and download it.
        """)
    with gr.Row():
        with gr.Column():
            with gr.Accordion(label="Generation Settings", open=True):
                seed = gr.Slider(0, MAX_SEED, label="Seed", value=1, step=1)
                randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
                gr.Markdown("Stage 1: Sparse Structure Generation")
                with gr.Row():
                    ss_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=7.5, step=0.1)
                    ss_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)
                gr.Markdown("Stage 2: Structured Latent Generation")
                with gr.Row():
                    slat_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=3.0, step=0.1)
                    slat_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)
            generate_btn = gr.Button("Generate")
        with gr.Column():
            video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
    
    # # Handlers
    demo.load(start_session)
    demo.unload(end_session)

    # ---------------------------
    # 原有交互逻辑（略）
    # ---------------------------
    input_image.upload(
        reset_image,
        [predictor, input_image],
        [predictor, original_image, message],
    )
    see_button.click(
        see_point, 
        inputs=[original_image, x_input, y_input], 
        outputs=[input_image]
    )
    add_button.click(
        add_point, 
        inputs=[x_input, y_input, visible_points_state], 
        outputs=[visible_points_state]
    )
    
    # ---------------------------
    # 新增的交互逻辑
    # ---------------------------
    clear_button.click(
        clear_all_points,
        inputs=[original_image],
        outputs=[input_image]
    )
    see_visible_button.click(
        see_visible_points,
        inputs=[input_image, visible_points_state],
        outputs=input_image
    )
    # 当 visible_points_state 或 occlusion_points_state 变化时，更新文本框和下拉菜单
    visible_points_state.change(
        update_all_points,
        inputs=[visible_points_state],
        outputs=[points_text, visible_points_dropdown]
    )
    delete_visible_button.click(
        delete_selected_visible,
        inputs=[input_image, visible_points_state, visible_points_dropdown],
        outputs=[input_image, visible_points_state, points_text, visible_points_dropdown]
    )

    # 生成mask的逻辑
    gen_vis_mask.click(
        segment_and_overlay,
        inputs=[original_image, visible_points_state, predictor],
        outputs=[visible_mask, visibility_mask]
    )
    add_vis_mask.click(
        add_mask,
        inputs=[visibility_mask, visibility_mask_list],
        outputs=[visibility_mask_list]
    )
    render_vis_mask.click(
        vis_mask,
        inputs=[original_image, visibility_mask_list],
        outputs=[visible_mask]
    )
    undo_vis_mask.click(
        delete_mask,
        inputs=[visibility_mask_list],
        outputs=[visibility_mask_list]
    )

    check_visible_input.click(
        check_combined_mask,
        inputs=[original_image, visibility_mask, visibility_mask_list, zoom_scale],
        outputs=[vis_input, occluded_mask]
    )
    

    # 3D Amodal Reconstruction
    # generate_btn.click(
    #     get_seed,
    #     inputs=[randomize_seed, seed],
    #     outputs=[seed],
    # ).then(
    #     image_to_3d,
    #     inputs=[vis_input, occluded_mask, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps],
    #     outputs=[output_buf, video_output],
    # )

    generate_btn.click(
        image_to_3d,
        inputs=[vis_input, occluded_mask, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps],
        outputs=[output_buf, video_output],
    )
    

# 启动 Gradio App
if __name__ == "__main__":
    pipeline = Amodal3RImageTo3DPipeline.from_pretrained("Sm0kyWu/Amodal3R")
    pipeline.cuda()
    try:
        pipeline.preprocess_image(Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8)))
    except:
        pass
    demo.launch()