File size: 4,900 Bytes
297a2c6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import gradio as gr
import sahi.utils
from sahi import AutoDetectionModel
import sahi.predict
import sahi.slicing
from PIL import Image
import numpy

IMAGE_SIZE = 640

# Images
sahi.utils.file.download_from_url(
    "https://user-images.githubusercontent.com/34196005/142730935-2ace3999-a47b-49bb-83e0-2bdd509f1c90.jpg",
    "apple_tree.jpg",
)
sahi.utils.file.download_from_url(
    "https://user-images.githubusercontent.com/34196005/142730936-1b397756-52e5-43be-a949-42ec0134d5d8.jpg",
    "highway.jpg",
)

sahi.utils.file.download_from_url(
    "https://user-images.githubusercontent.com/34196005/142742871-bf485f84-0355-43a3-be86-96b44e63c3a2.jpg",
    "highway2.jpg",
)

sahi.utils.file.download_from_url(
    "https://user-images.githubusercontent.com/34196005/142742872-1fefcc4d-d7e6-4c43-bbb7-6b5982f7e4ba.jpg",
    "highway3.jpg",
)


# Model
model = AutoDetectionModel.from_pretrained(
    model_type="yolov5", model_path="yolov5s6.pt", device="cpu", confidence_threshold=0.5, image_size=IMAGE_SIZE
)


def sahi_yolo_inference(
    image,
    slice_height=512,
    slice_width=512,
    overlap_height_ratio=0.2,
    overlap_width_ratio=0.2,
    postprocess_type="GREEDYNMM",
    postprocess_match_metric="IOS",
    postprocess_match_threshold=0.5,
    postprocess_class_agnostic=False,
):

    image_width, image_height = image.size
    sliced_bboxes = sahi.slicing.get_slice_bboxes(
        image_height,
        image_width,
        slice_height,
        slice_width,
        False,
        overlap_height_ratio,
        overlap_width_ratio,
    )
    if len(sliced_bboxes) > 60:
        raise ValueError(
            f"{len(sliced_bboxes)} slices are too much for huggingface spaces, try smaller slice size."
        )

    # standard inference
    prediction_result_1 = sahi.predict.get_prediction(
        image=image, detection_model=model
    )
    print(image)
    visual_result_1 = sahi.utils.cv.visualize_object_predictions(
        image=numpy.array(image),
        object_prediction_list=prediction_result_1.object_prediction_list,
    )
    output_1 = Image.fromarray(visual_result_1["image"])

    # sliced inference
    prediction_result_2 = sahi.predict.get_sliced_prediction(
        image=image,
        detection_model=model,
        slice_height=int(slice_height),
        slice_width=int(slice_width),
        overlap_height_ratio=overlap_height_ratio,
        overlap_width_ratio=overlap_width_ratio,
        postprocess_type=postprocess_type,
        postprocess_match_metric=postprocess_match_metric,
        postprocess_match_threshold=postprocess_match_threshold,
        postprocess_class_agnostic=postprocess_class_agnostic,
    )
    visual_result_2 = sahi.utils.cv.visualize_object_predictions(
        image=numpy.array(image),
        object_prediction_list=prediction_result_2.object_prediction_list,
    )

    output_2 = Image.fromarray(visual_result_2["image"])

    return output_1, output_2


inputs = [
    gr.inputs.Image(type="pil", label="Original Image"),
    gr.inputs.Number(default=512, label="slice_height"),
    gr.inputs.Number(default=512, label="slice_width"),
    gr.inputs.Number(default=0.2, label="overlap_height_ratio"),
    gr.inputs.Number(default=0.2, label="overlap_width_ratio"),
    gr.inputs.Dropdown(
        ["NMS", "GREEDYNMM"],
        type="value",
        default="GREEDYNMM",
        label="postprocess_type",
    ),
    gr.inputs.Dropdown(
        ["IOU", "IOS"], type="value", default="IOS", label="postprocess_type"
    ),
    gr.inputs.Number(default=0.5, label="postprocess_match_threshold"),
    gr.inputs.Checkbox(default=True, label="postprocess_class_agnostic"),
]

outputs = [
    gr.outputs.Image(type="pil", label="YOLOv5s"),
    gr.outputs.Image(type="pil", label="YOLOv5s + SAHI"),
]

title = "Small Object Detection with SAHI + YOLOv5"
description = "SAHI + YOLOv5 demo for small object detection. Upload an image or click an example image to use."
article = "<p style='text-align: center'>SAHI is a lightweight vision library for performing large scale object detection/ instance segmentation.. <a href='https://github.com/obss/sahi'>SAHI Github</a> | <a href='https://medium.com/codable/sahi-a-vision-library-for-performing-sliced-inference-on-large-images-small-objects-c8b086af3b80'>SAHI Blog</a> | <a href='https://github.com/fcakyon/yolov5-pip'>YOLOv5 Github</a> </p>"
examples = [
    ["apple_tree.jpg", 256, 256, 0.2, 0.2, "GREEDYNMM", "IOS", 0.5, True],
    ["highway.jpg", 256, 256, 0.2, 0.2, "GREEDYNMM", "IOS", 0.5, True],
    ["highway2.jpg", 512, 512, 0.2, 0.2, "GREEDYNMM", "IOS", 0.5, True],
    ["highway3.jpg", 512, 512, 0.2, 0.2, "GREEDYNMM", "IOS", 0.5, True],
]

gr.Interface(
    sahi_yolo_inference,
    inputs,
    outputs,
    title=title,
    description=description,
    article=article,
    examples=examples,
    theme="huggingface",
).launch(debug=True, enable_queue=True)