open vocabulary detection with Florence2 + masks with SAM2
Browse files- app.py +107 -57
- utils/florence.py +3 -0
- utils/modes.py +7 -0
- utils/sam.py +22 -0
app.py
CHANGED
|
@@ -1,15 +1,16 @@
|
|
| 1 |
from typing import Tuple, Optional
|
| 2 |
|
| 3 |
import gradio as gr
|
| 4 |
-
import numpy as np
|
| 5 |
import supervision as sv
|
| 6 |
import torch
|
| 7 |
from PIL import Image
|
| 8 |
|
| 9 |
from utils.florence import load_florence_model, run_florence_inference, \
|
| 10 |
FLORENCE_DETAILED_CAPTION_TASK, \
|
| 11 |
-
FLORENCE_CAPTION_TO_PHRASE_GROUNDING_TASK
|
| 12 |
-
from utils.
|
|
|
|
|
|
|
| 13 |
|
| 14 |
MARKDOWN = """
|
| 15 |
# Florence2 + SAM2 🔥
|
|
@@ -21,94 +22,122 @@ into masks.
|
|
| 21 |
"""
|
| 22 |
|
| 23 |
EXAMPLES = [
|
| 24 |
-
"https://media.roboflow.com/notebooks/examples/dog-2.jpeg",
|
| 25 |
-
"https://media.roboflow.com/notebooks/examples/dog-
|
| 26 |
-
"https://media.roboflow.com/notebooks/examples/dog-
|
|
|
|
|
|
|
| 27 |
]
|
| 28 |
|
| 29 |
DEVICE = torch.device("cuda")
|
| 30 |
-
|
| 31 |
FLORENCE_MODEL, FLORENCE_PROCESSOR = load_florence_model(device=DEVICE)
|
| 32 |
SAM_MODEL = load_sam_model(device=DEVICE)
|
| 33 |
BOX_ANNOTATOR = sv.BoxAnnotator(color_lookup=sv.ColorLookup.INDEX)
|
| 34 |
LABEL_ANNOTATOR = sv.LabelAnnotator(
|
| 35 |
color_lookup=sv.ColorLookup.INDEX,
|
| 36 |
text_position=sv.Position.CENTER_OF_MASS,
|
| 37 |
-
text_color=sv.Color.
|
| 38 |
border_radius=5
|
| 39 |
)
|
| 40 |
MASK_ANNOTATOR = sv.MaskAnnotator(color_lookup=sv.ColorLookup.INDEX)
|
| 41 |
|
| 42 |
|
| 43 |
-
def
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
|
|
|
| 48 |
|
| 49 |
-
_, result = run_florence_inference(
|
| 50 |
-
model=FLORENCE_MODEL,
|
| 51 |
-
processor=FLORENCE_PROCESSOR,
|
| 52 |
-
device=DEVICE,
|
| 53 |
-
image=image_input,
|
| 54 |
-
task=FLORENCE_DETAILED_CAPTION_TASK
|
| 55 |
-
)
|
| 56 |
-
caption = result[FLORENCE_DETAILED_CAPTION_TASK]
|
| 57 |
-
_, result = run_florence_inference(
|
| 58 |
-
model=FLORENCE_MODEL,
|
| 59 |
-
processor=FLORENCE_PROCESSOR,
|
| 60 |
-
device=DEVICE,
|
| 61 |
-
image=image_input,
|
| 62 |
-
task=FLORENCE_CAPTION_TO_PHRASE_GROUNDING_TASK,
|
| 63 |
-
text=caption
|
| 64 |
-
)
|
| 65 |
-
detections = sv.Detections.from_lmm(
|
| 66 |
-
lmm=sv.LMM.FLORENCE_2,
|
| 67 |
-
result=result,
|
| 68 |
-
resolution_wh=image_input.size
|
| 69 |
-
)
|
| 70 |
-
image = np.array(image_input.convert("RGB"))
|
| 71 |
-
SAM_MODEL.set_image(image)
|
| 72 |
-
mask, score, _ = SAM_MODEL.predict(box=detections.xyxy, multimask_output=False)
|
| 73 |
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
|
|
|
|
|
|
| 77 |
|
| 78 |
-
detections.mask = mask.astype(bool)
|
| 79 |
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 85 |
|
| 86 |
|
| 87 |
with gr.Blocks() as demo:
|
| 88 |
gr.Markdown(MARKDOWN)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 89 |
with gr.Row():
|
| 90 |
with gr.Column():
|
| 91 |
image_input_component = gr.Image(
|
| 92 |
type='pil', label='Upload image')
|
|
|
|
|
|
|
| 93 |
submit_button_component = gr.Button(value='Submit', variant='primary')
|
| 94 |
-
|
| 95 |
with gr.Column():
|
| 96 |
image_output_component = gr.Image(type='pil', label='Image output')
|
| 97 |
-
text_output_component = gr.Textbox(label='Caption output')
|
| 98 |
|
| 99 |
-
submit_button_component.click(
|
| 100 |
-
fn=process,
|
| 101 |
-
inputs=[image_input_component],
|
| 102 |
-
outputs=[
|
| 103 |
-
image_output_component,
|
| 104 |
-
text_output_component
|
| 105 |
-
]
|
| 106 |
-
)
|
| 107 |
with gr.Row():
|
| 108 |
gr.Examples(
|
| 109 |
fn=process,
|
| 110 |
examples=EXAMPLES,
|
| 111 |
-
inputs=[
|
|
|
|
|
|
|
|
|
|
|
|
|
| 112 |
outputs=[
|
| 113 |
image_output_component,
|
| 114 |
text_output_component
|
|
@@ -116,4 +145,25 @@ with gr.Blocks() as demo:
|
|
| 116 |
run_on_click=True
|
| 117 |
)
|
| 118 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 119 |
demo.launch(debug=False, show_error=True, max_threads=1)
|
|
|
|
| 1 |
from typing import Tuple, Optional
|
| 2 |
|
| 3 |
import gradio as gr
|
|
|
|
| 4 |
import supervision as sv
|
| 5 |
import torch
|
| 6 |
from PIL import Image
|
| 7 |
|
| 8 |
from utils.florence import load_florence_model, run_florence_inference, \
|
| 9 |
FLORENCE_DETAILED_CAPTION_TASK, \
|
| 10 |
+
FLORENCE_CAPTION_TO_PHRASE_GROUNDING_TASK, FLORENCE_OPEN_VOCABULARY_DETECTION_TASK
|
| 11 |
+
from utils.modes import INFERENCE_MODES, OPEN_VOCABULARY_DETECTION, \
|
| 12 |
+
CAPTION_GROUNDING_MASKS
|
| 13 |
+
from utils.sam import load_sam_model, run_sam_inference
|
| 14 |
|
| 15 |
MARKDOWN = """
|
| 16 |
# Florence2 + SAM2 🔥
|
|
|
|
| 22 |
"""
|
| 23 |
|
| 24 |
EXAMPLES = [
|
| 25 |
+
[OPEN_VOCABULARY_DETECTION, "https://media.roboflow.com/notebooks/examples/dog-2.jpeg", 'straw'],
|
| 26 |
+
[OPEN_VOCABULARY_DETECTION, "https://media.roboflow.com/notebooks/examples/dog-2.jpeg", 'napkin'],
|
| 27 |
+
[OPEN_VOCABULARY_DETECTION, "https://media.roboflow.com/notebooks/examples/dog-3.jpeg", 'tail'],
|
| 28 |
+
[CAPTION_GROUNDING_MASKS, "https://media.roboflow.com/notebooks/examples/dog-2.jpeg", None],
|
| 29 |
+
[CAPTION_GROUNDING_MASKS, "https://media.roboflow.com/notebooks/examples/dog-3.jpeg", None],
|
| 30 |
]
|
| 31 |
|
| 32 |
DEVICE = torch.device("cuda")
|
|
|
|
| 33 |
FLORENCE_MODEL, FLORENCE_PROCESSOR = load_florence_model(device=DEVICE)
|
| 34 |
SAM_MODEL = load_sam_model(device=DEVICE)
|
| 35 |
BOX_ANNOTATOR = sv.BoxAnnotator(color_lookup=sv.ColorLookup.INDEX)
|
| 36 |
LABEL_ANNOTATOR = sv.LabelAnnotator(
|
| 37 |
color_lookup=sv.ColorLookup.INDEX,
|
| 38 |
text_position=sv.Position.CENTER_OF_MASS,
|
| 39 |
+
text_color=sv.Color.from_hex("#FFFFFF"),
|
| 40 |
border_radius=5
|
| 41 |
)
|
| 42 |
MASK_ANNOTATOR = sv.MaskAnnotator(color_lookup=sv.ColorLookup.INDEX)
|
| 43 |
|
| 44 |
|
| 45 |
+
def annotate_image(image, detections):
|
| 46 |
+
output_image = image.copy()
|
| 47 |
+
output_image = MASK_ANNOTATOR.annotate(output_image, detections)
|
| 48 |
+
output_image = BOX_ANNOTATOR.annotate(output_image, detections)
|
| 49 |
+
output_image = LABEL_ANNOTATOR.annotate(output_image, detections)
|
| 50 |
+
return output_image
|
| 51 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
|
| 53 |
+
def on_mode_dropdown_change(text):
|
| 54 |
+
return [
|
| 55 |
+
gr.Textbox(visible=text == OPEN_VOCABULARY_DETECTION),
|
| 56 |
+
gr.Textbox(visible=text == CAPTION_GROUNDING_MASKS),
|
| 57 |
+
]
|
| 58 |
|
|
|
|
| 59 |
|
| 60 |
+
def process(
|
| 61 |
+
mode_dropdown, image_input, text_input
|
| 62 |
+
) -> Tuple[Optional[Image.Image], Optional[str]]:
|
| 63 |
+
if not image_input:
|
| 64 |
+
return None, None
|
| 65 |
+
|
| 66 |
+
if mode_dropdown == OPEN_VOCABULARY_DETECTION:
|
| 67 |
+
if not text_input:
|
| 68 |
+
return None, None
|
| 69 |
+
|
| 70 |
+
_, result = run_florence_inference(
|
| 71 |
+
model=FLORENCE_MODEL,
|
| 72 |
+
processor=FLORENCE_PROCESSOR,
|
| 73 |
+
device=DEVICE,
|
| 74 |
+
image=image_input,
|
| 75 |
+
task=FLORENCE_OPEN_VOCABULARY_DETECTION_TASK,
|
| 76 |
+
text=text_input
|
| 77 |
+
)
|
| 78 |
+
detections = sv.Detections.from_lmm(
|
| 79 |
+
lmm=sv.LMM.FLORENCE_2,
|
| 80 |
+
result=result,
|
| 81 |
+
resolution_wh=image_input.size
|
| 82 |
+
)
|
| 83 |
+
detections = run_sam_inference(SAM_MODEL, image_input, detections)
|
| 84 |
+
return annotate_image(image_input, detections), None
|
| 85 |
+
|
| 86 |
+
if mode_dropdown == CAPTION_GROUNDING_MASKS:
|
| 87 |
+
_, result = run_florence_inference(
|
| 88 |
+
model=FLORENCE_MODEL,
|
| 89 |
+
processor=FLORENCE_PROCESSOR,
|
| 90 |
+
device=DEVICE,
|
| 91 |
+
image=image_input,
|
| 92 |
+
task=FLORENCE_DETAILED_CAPTION_TASK
|
| 93 |
+
)
|
| 94 |
+
caption = result[FLORENCE_DETAILED_CAPTION_TASK]
|
| 95 |
+
_, result = run_florence_inference(
|
| 96 |
+
model=FLORENCE_MODEL,
|
| 97 |
+
processor=FLORENCE_PROCESSOR,
|
| 98 |
+
device=DEVICE,
|
| 99 |
+
image=image_input,
|
| 100 |
+
task=FLORENCE_CAPTION_TO_PHRASE_GROUNDING_TASK,
|
| 101 |
+
text=caption
|
| 102 |
+
)
|
| 103 |
+
detections = sv.Detections.from_lmm(
|
| 104 |
+
lmm=sv.LMM.FLORENCE_2,
|
| 105 |
+
result=result,
|
| 106 |
+
resolution_wh=image_input.size
|
| 107 |
+
)
|
| 108 |
+
detections = run_sam_inference(SAM_MODEL, image_input, detections)
|
| 109 |
+
return annotate_image(image_input, detections), caption
|
| 110 |
|
| 111 |
|
| 112 |
with gr.Blocks() as demo:
|
| 113 |
gr.Markdown(MARKDOWN)
|
| 114 |
+
mode_dropdown_component = gr.Dropdown(
|
| 115 |
+
choices=INFERENCE_MODES,
|
| 116 |
+
value=INFERENCE_MODES[0],
|
| 117 |
+
label="Mode",
|
| 118 |
+
info="Select a mode to use.",
|
| 119 |
+
interactive=True
|
| 120 |
+
)
|
| 121 |
with gr.Row():
|
| 122 |
with gr.Column():
|
| 123 |
image_input_component = gr.Image(
|
| 124 |
type='pil', label='Upload image')
|
| 125 |
+
text_input_component = gr.Textbox(
|
| 126 |
+
label='Text prompt')
|
| 127 |
submit_button_component = gr.Button(value='Submit', variant='primary')
|
|
|
|
| 128 |
with gr.Column():
|
| 129 |
image_output_component = gr.Image(type='pil', label='Image output')
|
| 130 |
+
text_output_component = gr.Textbox(label='Caption output', visible=False)
|
| 131 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 132 |
with gr.Row():
|
| 133 |
gr.Examples(
|
| 134 |
fn=process,
|
| 135 |
examples=EXAMPLES,
|
| 136 |
+
inputs=[
|
| 137 |
+
mode_dropdown_component,
|
| 138 |
+
image_input_component,
|
| 139 |
+
text_input_component
|
| 140 |
+
],
|
| 141 |
outputs=[
|
| 142 |
image_output_component,
|
| 143 |
text_output_component
|
|
|
|
| 145 |
run_on_click=True
|
| 146 |
)
|
| 147 |
|
| 148 |
+
submit_button_component.click(
|
| 149 |
+
fn=process,
|
| 150 |
+
inputs=[
|
| 151 |
+
mode_dropdown_component,
|
| 152 |
+
image_input_component,
|
| 153 |
+
text_input_component
|
| 154 |
+
],
|
| 155 |
+
outputs=[
|
| 156 |
+
image_output_component,
|
| 157 |
+
text_output_component
|
| 158 |
+
]
|
| 159 |
+
)
|
| 160 |
+
mode_dropdown_component.change(
|
| 161 |
+
on_mode_dropdown_change,
|
| 162 |
+
inputs=[mode_dropdown_component],
|
| 163 |
+
outputs=[
|
| 164 |
+
text_input_component,
|
| 165 |
+
text_output_component
|
| 166 |
+
]
|
| 167 |
+
)
|
| 168 |
+
|
| 169 |
demo.launch(debug=False, show_error=True, max_threads=1)
|
utils/florence.py
CHANGED
|
@@ -8,8 +8,11 @@ from transformers import AutoModelForCausalLM, AutoProcessor
|
|
| 8 |
from transformers.dynamic_module_utils import get_imports
|
| 9 |
|
| 10 |
FLORENCE_CHECKPOINT = "microsoft/Florence-2-large"
|
|
|
|
| 11 |
FLORENCE_DETAILED_CAPTION_TASK = '<MORE_DETAILED_CAPTION>'
|
| 12 |
FLORENCE_CAPTION_TO_PHRASE_GROUNDING_TASK = '<CAPTION_TO_PHRASE_GROUNDING>'
|
|
|
|
|
|
|
| 13 |
|
| 14 |
|
| 15 |
def fixed_get_imports(filename: Union[str, os.PathLike]) -> list[str]:
|
|
|
|
| 8 |
from transformers.dynamic_module_utils import get_imports
|
| 9 |
|
| 10 |
FLORENCE_CHECKPOINT = "microsoft/Florence-2-large"
|
| 11 |
+
FLORENCE_OBJECT_DETECTION_TASK = '<OD>'
|
| 12 |
FLORENCE_DETAILED_CAPTION_TASK = '<MORE_DETAILED_CAPTION>'
|
| 13 |
FLORENCE_CAPTION_TO_PHRASE_GROUNDING_TASK = '<CAPTION_TO_PHRASE_GROUNDING>'
|
| 14 |
+
FLORENCE_OPEN_VOCABULARY_DETECTION_TASK = '<OPEN_VOCABULARY_DETECTION>'
|
| 15 |
+
FLORENCE_DENSE_REGION_CAPTION_TASK = '<DENSE_REGION_CAPTION>'
|
| 16 |
|
| 17 |
|
| 18 |
def fixed_get_imports(filename: Union[str, os.PathLike]) -> list[str]:
|
utils/modes.py
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
OPEN_VOCABULARY_DETECTION = "open vocabulary detection + masks"
|
| 2 |
+
CAPTION_GROUNDING_MASKS = "caption + grounding + masks"
|
| 3 |
+
|
| 4 |
+
INFERENCE_MODES = [
|
| 5 |
+
OPEN_VOCABULARY_DETECTION,
|
| 6 |
+
CAPTION_GROUNDING_MASKS
|
| 7 |
+
]
|
utils/sam.py
CHANGED
|
@@ -1,4 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import torch
|
|
|
|
| 2 |
from sam2.build_sam import build_sam2
|
| 3 |
from sam2.sam2_image_predictor import SAM2ImagePredictor
|
| 4 |
|
|
@@ -13,3 +18,20 @@ def load_sam_model(
|
|
| 13 |
) -> SAM2ImagePredictor:
|
| 14 |
model = build_sam2(config, checkpoint, device=device)
|
| 15 |
return SAM2ImagePredictor(sam_model=model)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Any
|
| 2 |
+
|
| 3 |
+
import numpy as np
|
| 4 |
+
import supervision as sv
|
| 5 |
import torch
|
| 6 |
+
from PIL import Image
|
| 7 |
from sam2.build_sam import build_sam2
|
| 8 |
from sam2.sam2_image_predictor import SAM2ImagePredictor
|
| 9 |
|
|
|
|
| 18 |
) -> SAM2ImagePredictor:
|
| 19 |
model = build_sam2(config, checkpoint, device=device)
|
| 20 |
return SAM2ImagePredictor(sam_model=model)
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
def run_sam_inference(
|
| 24 |
+
model: Any,
|
| 25 |
+
image: Image,
|
| 26 |
+
detections: sv.Detections
|
| 27 |
+
) -> sv.Detections:
|
| 28 |
+
image = np.array(image.convert("RGB"))
|
| 29 |
+
model.set_image(image)
|
| 30 |
+
mask, score, _ = model.predict(box=detections.xyxy, multimask_output=False)
|
| 31 |
+
|
| 32 |
+
# dirty fix; remove this later
|
| 33 |
+
if len(mask.shape) == 4:
|
| 34 |
+
mask = np.squeeze(mask)
|
| 35 |
+
|
| 36 |
+
detections.mask = mask.astype(bool)
|
| 37 |
+
return detections
|