File size: 7,574 Bytes
a13210a
 
 
 
 
 
 
 
 
 
42404e5
b036ed3
a13210a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a3c9048
a13210a
 
 
 
 
 
 
 
 
 
 
 
 
a3c9048
 
a13210a
 
 
a3c9048
a13210a
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
"""Gradio layout and bundled examples for the RSReasonNet demo."""

import json
from pathlib import Path

import gradio as gr


EXAMPLE_DIR = Path(__file__).resolve().parent / "example"
EXAMPLE_TITLES = {
    "03261.jpg": "Ship · Maritime logistics",
    "04077.jpg": "Storage tank · Industrial infrastructure",
    "02026.jpg": "Tennis court · Everyday language",
}


def load_examples():
    """Keep each image paired with its original annotated instruction."""
    with (EXAMPLE_DIR / "annotations.json").open(encoding="utf-8") as handle:
        annotations = json.load(handle)
    examples = []
    for annotation in annotations:
        image_path = EXAMPLE_DIR / annotation["id"]
        if not image_path.is_file():
            raise FileNotFoundError(f"Example image is missing: {image_path.name}")
        examples.append((str(image_path), annotation["sentence"]))
    return examples


CSS = """
.gradio-container { max-width: 1120px !important; margin: auto; }
#rs-header { text-align: center; padding: 22px 16px 18px; }
#rs-header .eyebrow {
    display: inline-block; padding: 5px 12px; border-radius: 999px;
    border: 1px solid var(--border-color-primary);
    color: var(--body-text-color-subdued); font-size: 12px; letter-spacing: .05em;
}
#rs-header h1 {
    margin: 12px 0 8px; font-size: clamp(30px, 5vw, 42px);
    font-weight: 700; letter-spacing: -.04em; color: var(--body-text-color);
}
#rs-header p {
    margin: 0 auto; max-width: 640px; color: var(--body-text-color-subdued);
    font-size: 15px; line-height: 1.6;
}
.rs-panel { padding: 18px !important; border-radius: 16px !important; }
.rs-section h2 { font-size: 18px !important; margin: 0 0 4px !important; }
.rs-section p { color: var(--body-text-color-subdued); font-size: 13px; }
#rs-run { min-height: 44px; font-weight: 600; }
#rs-examples-heading { margin-top: 12px; }
#rs-examples { border-radius: 14px; }
#rs-examples .grid-wrap { height: auto; min-height: 0; max-height: none; }
#rs-examples .grid-container {
    grid-template-columns: repeat(3, minmax(0, 1fr));
    grid-template-rows: none;
}
#rs-examples .thumbnail-item {
    cursor: pointer; aspect-ratio: auto; height: 210px;
}
#rs-examples .thumbnail-item:hover { border-color: var(--color-accent); }
#rs-examples .caption-label { font-size: 13px; }
#rs-note { color: var(--body-text-color-subdued); font-size: 12px; }
@media (max-width: 640px) {
    #rs-header { padding: 14px 4px; }
    .rs-panel { padding: 12px !important; }
    #rs-examples .grid-container { grid-template-columns: minmax(0, 1fr); }
}
"""


def build_demo(predict):
    examples = load_examples()
    theme = gr.themes.Soft(
        primary_hue="blue", secondary_hue="slate", neutral_hue="slate",
        font=["system-ui", "sans-serif"],
    )
    with gr.Blocks(
        title="RSReasonNet · Reasoning Segmentation",
        theme=theme,
        css=CSS,
        analytics_enabled=False,
    ) as demo:
        gr.HTML("""
            <header id="rs-header">
                <span class="eyebrow">🛰️ REMOTE SENSING · REASONING SEGMENTATION</span>
                <h1>RSReasonNet</h1>
                <p>Describe what you are looking for. Explore how language guides
                segmentation in satellite imagery.</p>
            </header>
        """)

        with gr.Row(equal_height=True):
            with gr.Column(min_width=300, variant="panel", elem_classes="rs-panel"):
                gr.Markdown(
                    "## 1. Image & instruction\nUpload an image or choose an example below.",
                    elem_classes="rs-section",
                )
                input_image = gr.Image(
                    value=examples[0][0], type="filepath", sources=["upload"],
                    label="Satellite image", height=320, elem_id="rs-input-image",
                    show_share_button=False,
                )
                instruction = gr.Textbox(
                    value=examples[0][1], label="What should the model find?",
                    placeholder="Describe an object, its purpose, or a clue in English…",
                    lines=3, max_lines=5,
                    info="Use an English instruction (up to 1,024 characters).",
                )
                with gr.Row():
                    run_button = gr.Button(
                        "Run segmentation", variant="primary", scale=3, elem_id="rs-run",
                    )
                    clear_button = gr.ClearButton(value="Clear", scale=1)

            with gr.Column(min_width=300, variant="panel", elem_classes="rs-panel"):
                gr.Markdown(
                    "## 2. Segmentation & response\nRun the model to see the region it identifies.",
                    elem_classes="rs-section",
                )
                output_image = gr.Image(
                    type="numpy", label="Segmentation overlay", height=320,
                    interactive=False, show_share_button=False, elem_id="rs-output-image",
                )
                text_output = gr.Textbox(
                    label="Model response", lines=3, max_lines=6,
                    interactive=False, show_copy_button=True,
                    placeholder="The model's text response will appear here.",
                )
                gr.Markdown(
                    "🔴 **Red overlay** marks the predicted region. "
                    "Some instructions may return text without a mask.",
                    elem_id="rs-note",
                )

        gr.Markdown(
            "## Try an example\nClick a thumbnail to load its image and original instruction, "
            "then select **Run segmentation**.",
            elem_id="rs-examples-heading", elem_classes="rs-section",
        )
        gallery = gr.Gallery(
            value=[(path, EXAMPLE_TITLES.get(Path(path).name, Path(path).stem))
                   for path, _ in examples],
            label="Example images", show_label=False,
            columns=3, rows=1,
            object_fit="contain", allow_preview=False, interactive=False,
            show_share_button=False, show_download_button=False,
            show_fullscreen_button=False, elem_id="rs-examples",
        )

        # Keep the existing API argument order: instruction first, image second.
        run_event = gr.on(
            triggers=[run_button.click, instruction.submit], fn=predict,
            inputs=[instruction, input_image], outputs=[output_image, text_output],
            api_name="predict", concurrency_limit=1, concurrency_id="segmentation",
            show_progress="full",
        )

        def select_example(event: gr.SelectData):
            image_path, prompt = examples[event.index]
            # Allow the same card to be selected again after clearing or editing.
            return image_path, prompt, None, "", gr.Gallery(selected_index=None)

        gallery.select(
            select_example, inputs=None,
            outputs=[input_image, instruction, output_image, text_output, gallery],
            queue=False, api_name=False, cancels=[run_event],
        )
        clear_button.add([input_image, instruction, output_image, text_output])
        clear_button.click(fn=None, cancels=[run_event], queue=False, api_name=False)

        # A result belongs to its input pair; remove it when that pair is edited.
        gr.on(
            triggers=[input_image.input, instruction.input], fn=lambda: (None, ""),
            inputs=None, outputs=[output_image, text_output], queue=False,
            api_name=False, cancels=[run_event],
        )

    return demo