app.py

import glob
import gradio as gr
import numpy as np
from PIL import Image
from transformers import MaskFormerFeatureExtractor, MaskFormerForInstanceSegmentation


# feature_extractor = MaskFormerFeatureExtractor.from_pretrained("facebook/maskformer-swin-tiny-coco")
# model = MaskFormerForInstanceSegmentation.from_pretrained("facebook/maskformer-swin-tiny-coco")
feature_extractor = MaskFormerFeatureExtractor.from_pretrained("facebook/maskformer-swin-large-coco")
model = MaskFormerForInstanceSegmentation.from_pretrained("facebook/maskformer-swin-large-coco")

example_images = sorted(glob.glob('examples/map*.jpg'))

def visualize_instance_seg_mask(img_in, mask, id2label):
    img_out = np.zeros((mask.shape[0], mask.shape[1], 3))
    image_total_pixels = mask.shape[0] * mask.shape[1]
    label_ids = np.unique(mask)
    vegetation_labels = ["tree-merged", "grass-merged"]

    def get_color(id):
        id_color = (np.random.randint(0, 2), np.random.randint(0, 4), np.random.randint(0, 256))
        if id2label[id] in vegetation_labels:
            id_color = (0, 140, 0)
        return id_color

    id2color = {id: get_color(id) for id in label_ids}
    id2count = {id: 0 for id in label_ids}

    for i in range(img_out.shape[0]):
      for j in range(img_out.shape[1]):
        img_out[i, j, :] = id2color[mask[i, j]]
        id2count[mask[i, j]] = id2count[mask[i, j]] + 1

    image_res = (0.5 * img_in + 0.5 * img_out).astype(np.uint8)

    vegetation_count = sum([id2count[id] for id in label_ids if id2label[id] in vegetation_labels])

    dataframe_vegetation_items = [[
        f"{id2label[id]}",
        f"{(100 * id2count[id] / image_total_pixels):.2f} %",
        f"{np.sqrt(id2count[id] / image_total_pixels):.2f} m"
        ] for id in label_ids if id2label[id] in vegetation_labels]
    dataframe_all_items = [[
        f"{id2label[id]}",
        f"{(100 * id2count[id] / image_total_pixels):.2f} %",
        f"{np.sqrt(id2count[id] / image_total_pixels):.2f} m"
        ] for id in label_ids]
    dataframe_vegetation_total = [[
        f"vegetation",
        f"{(100 * vegetation_count / image_total_pixels):.2f} %",
        f"{np.sqrt(vegetation_count / image_total_pixels):.2f} m"]]

    dataframe = dataframe_vegetation_total
    if len(dataframe) < 1:
        dataframe = [[
            f"",
            f"{(0):.2f} %",
            f"{(0):.2f} m"
        ]]

    return image_res, dataframe


def query_image(image_path):
    img = np.array(Image.open(image_path))
    img_size = (img.shape[0], img.shape[1])
    inputs = feature_extractor(images=img, return_tensors="pt")
    outputs = model(**inputs)
    results = feature_extractor.post_process_semantic_segmentation(outputs=outputs, target_sizes=[img_size])[0]
    mask_img, dataframe = visualize_instance_seg_mask(img, results.numpy(), model.config.id2label)
    return mask_img, dataframe


demo = gr.Interface(
    query_image,
    inputs=[gr.Image(type="filepath", label="Input Image")],
    outputs=[
        gr.Image(label="Vegetation"),
        gr.DataFrame(label="Info", headers=["Object Label", "Pixel Percent", "Square Length"])
    ],
    title="Maskformer (large-coco)",
    allow_flagging="never",
    analytics_enabled=None,
    examples=example_images,
    cache_examples=True
)

demo.launch(show_api=False)