matte_anything.py

import os
import cv2
import torch
import numpy as np
import gradio as gr
from PIL import Image
from torchvision.ops import box_convert
from detectron2.config import LazyConfig, instantiate
from detectron2.checkpoint import DetectionCheckpointer
from segment_anything import sam_model_registry, SamPredictor
import groundingdino.datasets.transforms as T
from groundingdino.util.inference import load_model as dino_load_model, predict as dino_predict, annotate as dino_annotate

models = {
	'vit_h': './pretrained/sam_vit_h_4b8939.pth',
    'vit_b': './pretrained/sam_vit_b_01ec64.pth'
}

vitmatte_models = {
	'vit_b': './pretrained/ViTMatte_B_DIS.pth',
}

vitmatte_config = {
	'vit_b': './configs/matte_anything.py',
}

grounding_dino = {
    'config': './GroundingDINO/groundingdino/config/GroundingDINO_SwinT_OGC.py',
    'weight': './pretrained/groundingdino_swint_ogc.pth'
}

def generate_checkerboard_image(height, width, num_squares):
    num_squares_h = num_squares
    square_size_h = height // num_squares_h
    square_size_w = square_size_h
    num_squares_w = width // square_size_w


    new_height = num_squares_h * square_size_h
    new_width = num_squares_w * square_size_w
    image = np.zeros((new_height, new_width), dtype=np.uint8)

    for i in range(num_squares_h):
        for j in range(num_squares_w):
            start_x = j * square_size_w
            start_y = i * square_size_h
            color = 255 if (i + j) % 2 == 0 else 200
            image[start_y:start_y + square_size_h, start_x:start_x + square_size_w] = color

    image = cv2.resize(image, (width, height))
    image = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB)

    return image


def init_segment_anything(model_type):

    #Initialize the segmenting anything with model_type in ['vit_b', 'vit_l', 'vit_h']


    sam = sam_model_registry[model_type](checkpoint=models[model_type]).to(device)
    predictor = SamPredictor(sam)

    return predictor

def init_vitmatte(model_type):

    #Initialize the vitmatte with model_type in ['vit_s', 'vit_b']

    cfg = LazyConfig.load(vitmatte_config[model_type])
    vitmatte = instantiate(cfg.model)
    vitmatte.to(device)
    vitmatte.eval()
    DetectionCheckpointer(vitmatte).load(vitmatte_models[model_type])

    return vitmatte

def generate_trimap(mask, erode_kernel_size=10, dilate_kernel_size=10):
    erode_kernel = np.ones((erode_kernel_size, erode_kernel_size), np.uint8)
    dilate_kernel = np.ones((dilate_kernel_size, dilate_kernel_size), np.uint8)
    eroded = cv2.erode(mask, erode_kernel, iterations=5)
    dilated = cv2.dilate(mask, dilate_kernel, iterations=5)
    trimap = np.zeros_like(mask)
    trimap[dilated==255] = 128
    trimap[eroded==255] = 255
    return trimap

# user click the image to get points, and show the points on the image
def get_point(img, sel_pix, point_type, evt: gr.SelectData):
    if point_type == 'foreground_point':
        sel_pix.append((evt.index, 1))   # append the foreground_point
    elif point_type == 'background_point':
        sel_pix.append((evt.index, 0))    # append the background_point
    else:
        sel_pix.append((evt.index, 1))    # default foreground_point
    # draw points
    for point, label in sel_pix:
        cv2.drawMarker(img, point, colors[label], markerType=markers[label], markerSize=20, thickness=5)
    if img[..., 0][0, 0] == img[..., 2][0, 0]:  # BGR to RGB
        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    return img if isinstance(img, np.ndarray) else np.array(img)

# undo the selected point
def undo_points(orig_img, sel_pix):
    temp = orig_img.copy()
    # draw points
    if len(sel_pix) != 0:
        sel_pix.pop()
        for point, label in sel_pix:
            cv2.drawMarker(temp, point, colors[label], markerType=markers[label], markerSize=20, thickness=5)
    if temp[..., 0][0, 0] == temp[..., 2][0, 0]:  # BGR to RGB
        temp = cv2.cvtColor(temp, cv2.COLOR_BGR2RGB)
    return temp if isinstance(temp, np.ndarray) else np.array(temp)

# once user upload an image, the original image is stored in `original_image`
def store_img(img):
    return img, []  # when new image is uploaded, `selected_points` should be empty

def convert_pixels(gray_image, boxes):
    converted_image = np.copy(gray_image)

    for box in boxes:
        x1, y1, x2, y2 = box
        x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2)
        converted_image[y1:y2, x1:x2][converted_image[y1:y2, x1:x2] == 1] = 0.5

    return converted_image

if __name__ == "__main__":
    device = 'cuda'
    sam_model = 'vit_h'
    vitmatte_model = 'vit_b'

    colors = [(255, 0, 0), (0, 255, 0)]
    markers = [1, 5]

    print('Initializing models... Please wait...')

    predictor = init_segment_anything(sam_model)
    vitmatte = init_vitmatte(vitmatte_model)
    grounding_dino = dino_load_model(grounding_dino['config'], grounding_dino['weight'])

    def run_inference(input_x, selected_points, erode_kernel_size, dilate_kernel_size, fg_box_threshold, fg_text_threshold, fg_caption, tr_box_threshold, tr_text_threshold, tr_caption):
        predictor.set_image(input_x)

        dino_transform = T.Compose(
        [
            T.RandomResize([800], max_size=1333),
            T.ToTensor(),
            T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
        ])
        image_transformed, _ = dino_transform(Image.fromarray(input_x), None)

        if len(selected_points) != 0:
            points = torch.Tensor([p for p, _ in selected_points]).to(device).unsqueeze(1)
            labels = torch.Tensor([int(l) for _, l in selected_points]).to(device).unsqueeze(1)
            transformed_points = predictor.transform.apply_coords_torch(points, input_x.shape[:2])
            print(points.size(), transformed_points.size(), labels.size(), input_x.shape, points)
            point_coords=transformed_points.permute(1, 0, 2)
            point_labels=labels.permute(1, 0)
        else:
            transformed_points, labels = None, None
            point_coords, point_labels = None, None

        if fg_caption is not None and fg_caption != "":
            fg_boxes, logits, phrases = dino_predict(
                model=grounding_dino,
                image=image_transformed,
                caption=fg_caption,
                box_threshold=fg_box_threshold,
                text_threshold=fg_text_threshold
                )
            print(logits, phrases)
            if fg_boxes.shape[0] == 0:
                # no fg object detected
                transformed_boxes = None
            else:
                h, w, _ = input_x.shape
                fg_boxes = torch.Tensor(fg_boxes).to(device)
                fg_boxes = fg_boxes * torch.Tensor([w, h, w, h]).to(device)
                fg_boxes = box_convert(boxes=fg_boxes, in_fmt="cxcywh", out_fmt="xyxy")
                transformed_boxes = predictor.transform.apply_boxes_torch(fg_boxes, input_x.shape[:2])

        else:
            transformed_boxes = None

        print(transformed_boxes)
        # predict segmentation according to the boxes or points
        masks, scores, logits = predictor.predict_torch(
            point_coords=point_coords,
            point_labels=point_labels,
            boxes=transformed_boxes,
            multimask_output=False,
        )
        print(masks.shape, logits.shape)
        masks = masks.cpu().detach().numpy()
        mask_all = np.ones((input_x.shape[0], input_x.shape[1], 3))
        for ann in masks:
            color_mask = np.random.random((1, 3)).tolist()[0]
            for i in range(3):
                mask_all[ann[0] == True, i] = color_mask[i]
        img = input_x / 255 * 0.3 + mask_all * 0.7

        # generate alpha matte
        torch.cuda.empty_cache()
        mask = np.zeros(masks[0][0].shape)
        print("Mask shape ", mask.shape, masks.shape[0])
        for i in range(masks.shape[0]):
            mask[masks[i][0] == True] = 255
        mask = mask.astype(np.uint8)
        print("Trimap in values: ", np.unique(mask))
        trimap = generate_trimap(mask, erode_kernel_size, dilate_kernel_size).astype(np.float32)
        print("Trimap out values: ", np.unique(trimap))
        trimap[trimap==128] = 0.5
        trimap[trimap==255] = 1

        boxes, logits, phrases = dino_predict(
            model=grounding_dino,
            image=image_transformed,
            caption=tr_caption,
            box_threshold=tr_box_threshold,
            text_threshold=tr_text_threshold
            )
        annotated_frame = dino_annotate(image_source=input_x, boxes=boxes, logits=logits, phrases=phrases)
        # 把annotated_frame的改成RGB
        annotated_frame = cv2.cvtColor(annotated_frame, cv2.COLOR_BGR2RGB)

        if boxes.shape[0] == 0:
            # no transparent object detected
            pass
        else:
            h, w, _ = input_x.shape
            boxes = boxes * torch.Tensor([w, h, w, h])
            xyxy = box_convert(boxes=boxes, in_fmt="cxcywh", out_fmt="xyxy").numpy()
            trimap = convert_pixels(trimap, xyxy)

        input = {
            "image": torch.from_numpy(input_x).permute(2, 0, 1).unsqueeze(0)/255,
            "trimap": torch.from_numpy(trimap).unsqueeze(0).unsqueeze(0),
        }

        torch.cuda.empty_cache()
        alpha = vitmatte(input)['phas'].flatten(0,2)
        alpha = alpha.detach().cpu().numpy()

        # get a green background
        background = generate_checkerboard_image(input_x.shape[0], input_x.shape[1], 8)

        # calculate foreground with alpha blending
        foreground_alpha = input_x * np.expand_dims(alpha, axis=2).repeat(3,2)/255 + background * (1 - np.expand_dims(alpha, axis=2).repeat(3,2))/255

        # calculate foreground with mask
        foreground_mask = input_x * np.expand_dims(mask/255, axis=2).repeat(3,2)/255 + background * (1 - np.expand_dims(mask/255, axis=2).repeat(3,2))/255

        foreground_alpha[foreground_alpha>1] = 1
        foreground_mask[foreground_mask>1] = 1

        # return img, mask_all
        trimap[trimap==1] == 0.999

        # new background

        background_1 = cv2.imread('figs/sea.jpg')
        background_2 = cv2.imread('figs/forest.jpg')
        background_3 = cv2.imread('figs/sunny.jpg')

        background_1 = cv2.resize(background_1, (input_x.shape[1], input_x.shape[0]))
        background_2 = cv2.resize(background_2, (input_x.shape[1], input_x.shape[0]))
        background_3 = cv2.resize(background_3, (input_x.shape[1], input_x.shape[0]))

        # to RGB
        background_1 = cv2.cvtColor(background_1, cv2.COLOR_BGR2RGB)
        background_2 = cv2.cvtColor(background_2, cv2.COLOR_BGR2RGB)
        background_3 = cv2.cvtColor(background_3, cv2.COLOR_BGR2RGB)

        # use alpha blending
        new_bg_1 = input_x * np.expand_dims(alpha, axis=2).repeat(3,2)/255 + background_1 * (1 - np.expand_dims(alpha, axis=2).repeat(3,2))/255
        new_bg_2 = input_x * np.expand_dims(alpha, axis=2).repeat(3,2)/255 + background_2 * (1 - np.expand_dims(alpha, axis=2).repeat(3,2))/255
        new_bg_3 = input_x * np.expand_dims(alpha, axis=2).repeat(3,2)/255 + background_3 * (1 - np.expand_dims(alpha, axis=2).repeat(3,2))/255

        return  mask, alpha,  foreground_mask, foreground_alpha, new_bg_1, new_bg_2, new_bg_3

    
    with gr.Blocks() as demo:
        gr.Markdown(

            "# <center>Matte Anything🐒 !"

        )
        with gr.Row().style(equal_height=True):
            with gr.Column():
                # input image
                original_image = gr.State(value=None)   # store original image without points, default None
                input_image = gr.Image(type="numpy")


                # show the image with mask
                with gr.Tab(label='SAM Mask'):
                    mask = gr.Image(type='numpy')
                # with gr.Tab(label='Trimap'):
                #     trimap = gr.Image(type='numpy')
                with gr.Tab(label='Alpha Matte'):
                    alpha = gr.Image(type='numpy')
                # show only mask
                with gr.Tab(label='Foreground by SAM Mask'):
                    foreground_by_sam_mask = gr.Image(type='numpy')
                with gr.Tab(label='Refined by ViTMatte'):
                    refined_by_vitmatte = gr.Image(type='numpy')
                # with gr.Tab(label='Transparency Detection'):
                #     transparency = gr.Image(type='numpy')
                with gr.Tab(label='New Background 1'):
                    new_bg_1 = gr.Image(type='numpy')
                with gr.Tab(label='New Background 2'):
                    new_bg_2 = gr.Image(type='numpy')
                with gr.Tab(label='New Background 3'):
                    new_bg_3 = gr.Image(type='numpy')


            with gr.Column():
                # point prompt
                with gr.Column():
                    selected_points = gr.State([])      # store points
                    with gr.Row():
                        radio = gr.Radio(['foreground_point', 'background_point'], label='point labels')
                        undo_button = gr.Button('Remove Point')
                gr.Markdown("Please select points OR provide text input. Both together can result in error due to any mismatch.")
                fg_caption = gr.inputs.Textbox(lines=1, default="", label="foreground input text")

                # run button
                button = gr.Button("Start!")

                with gr.Box():
                    gr.Markdown("Trimap Settings")
                    erode_kernel_size = gr.inputs.Slider(minimum=1, maximum=30, step=1, default=10, label="erode_kernel_size")
                    dilate_kernel_size = gr.inputs.Slider(minimum=1, maximum=30, step=1, default=10, label="dilate_kernel_size")
                
                with gr.Box():
                    gr.Markdown("Input Text Settings")
                    fg_box_threshold = gr.inputs.Slider(minimum=0.0, maximum=1.0, step=0.001, default=0.25, label="foreground_box_threshold")
                    fg_text_threshold = gr.inputs.Slider(minimum=0.0, maximum=1.0, step=0.001, default=0.25, label="foreground_text_threshold")

                with gr.Box():
                    gr.Markdown("Transparency Settings")
                    tr_caption = gr.inputs.Textbox(lines=1, default="glass.lens.crystal.diamond.bubble.bulb.web.grid", label="transparency input text")
                    tr_box_threshold = gr.inputs.Slider(minimum=0.0, maximum=1.0, step=0.005, default=0.5, label="transparency_box_threshold")
                    tr_text_threshold = gr.inputs.Slider(minimum=0.0, maximum=1.0, step=0.005, default=0.25, label="transparency_text_threshold")



        input_image.upload(
            store_img,
            [input_image],
            [original_image, selected_points]
        )
        input_image.select(
            get_point,
            [input_image, selected_points, radio],
            [input_image],
        )
        undo_button.click(
            undo_points,
            [original_image, selected_points],
            [input_image]
        )
        button.click(run_inference, inputs=[original_image, selected_points, erode_kernel_size, dilate_kernel_size, fg_box_threshold, fg_text_threshold, fg_caption, tr_box_threshold, tr_text_threshold, tr_caption], outputs=[mask, alpha,  \
                                            foreground_by_sam_mask, refined_by_vitmatte, new_bg_1, new_bg_2, new_bg_3])

        with gr.Row():
            with gr.Column():
                background_image = gr.State(value=None)

    demo.launch(share=True)