Initial commit with model files

Dockerfile

@@ -0,0 +1,35 @@
+# Use Python 3.9 slim base image
+FROM python:3.9-slim
+
+# Install system dependencies
+RUN apt-get update && apt-get install -y \
+    libgl1-mesa-glx \
+    libglib2.0-0 \
+    wget \
+    git \
+    && rm -rf /var/lib/apt/lists/*
+
+# Set working directory
+WORKDIR /app
+
+# Copy requirements and install dependencies
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+# Create directories for models and temporary files with proper permissions
+RUN mkdir -p /app/weights/icon_detect /app/weights/icon_caption_florence /app/imgs \
+    && chown -R 1000:1000 /app
+
+# Copy application code
+COPY main.py utils.py download_models.py ./
+
+# Download models during build
+RUN python download_models.py
+
+# Set up user for HF Spaces
+RUN useradd -m -u 1000 user
+USER user
+ENV PATH="/home/user/.local/bin:$PATH"
+
+# Run the application
+CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "7860"]

app.py

@@ -0,0 +1,133 @@
+from fastapi import FastAPI, File, UploadFile, HTTPException
+from fastapi.responses import JSONResponse
+from pydantic import BaseModel
+from typing import Optional
+import base64
+import io
+from PIL import Image
+import torch
+import numpy as np
+import os
+
+# Existing imports
+import numpy as np
+import torch
+from PIL import Image
+import io
+
+from utils import (
+    check_ocr_box,
+    get_yolo_model,
+    get_caption_model_processor,
+    get_som_labeled_img,
+)
+import torch
+
+# yolo_model = get_yolo_model(model_path='/data/icon_detect/best.pt')
+# caption_model_processor = get_caption_model_processor(model_name="florence2", model_name_or_path="/data/icon_caption_florence")
+
+from ultralytics import YOLO
+
+# if not os.path.exists("/data/icon_detect"):
+#     os.makedirs("/data/icon_detect")
+
+try:
+    yolo_model = torch.load("weights/icon_detect/best.pt", map_location="cuda", weights_only=False)["model"]
+    yolo_model = yolo_model.to("cuda")
+except:
+    yolo_model = torch.load("weights/icon_detect/best.pt", map_location="cpu", weights_only=False)["model"]
+
+from transformers import AutoProcessor, AutoModelForCausalLM
+
+processor = AutoProcessor.from_pretrained(
+    "microsoft/Florence-2-base", trust_remote_code=True
+)
+
+try:
+    model = AutoModelForCausalLM.from_pretrained(
+        "weights/icon_caption_florence",
+        torch_dtype=torch.float16,
+        trust_remote_code=True,
+    ).to("cuda")
+except:
+    model = AutoModelForCausalLM.from_pretrained(
+        "weights/icon_caption_florence",
+        torch_dtype=torch.float16,
+        trust_remote_code=True,
+    )
+caption_model_processor = {"processor": processor, "model": model}
+print("finish loading model!!!")
+
+app = FastAPI()
+
+
+class ProcessResponse(BaseModel):
+    image: str  # Base64 encoded image
+    parsed_content_list: str
+    label_coordinates: str
+
+
+def process(
+    image_input: Image.Image, box_threshold: float, iou_threshold: float
+) -> ProcessResponse:
+    image_save_path = "imgs/saved_image_demo.png"
+    image_input.save(image_save_path)
+    image = Image.open(image_save_path)
+    box_overlay_ratio = image.size[0] / 3200
+    draw_bbox_config = {
+        "text_scale": 0.8 * box_overlay_ratio,
+        "text_thickness": max(int(2 * box_overlay_ratio), 1),
+        "text_padding": max(int(3 * box_overlay_ratio), 1),
+        "thickness": max(int(3 * box_overlay_ratio), 1),
+    }
+
+    ocr_bbox_rslt, is_goal_filtered = check_ocr_box(
+        image_save_path,
+        display_img=False,
+        output_bb_format="xyxy",
+        goal_filtering=None,
+        easyocr_args={"paragraph": False, "text_threshold": 0.9},
+        use_paddleocr=True,
+    )
+    text, ocr_bbox = ocr_bbox_rslt
+    dino_labled_img, label_coordinates, parsed_content_list = get_som_labeled_img(
+        image_save_path,
+        yolo_model,
+        BOX_TRESHOLD=box_threshold,
+        output_coord_in_ratio=True,
+        ocr_bbox=ocr_bbox,
+        draw_bbox_config=draw_bbox_config,
+        caption_model_processor=caption_model_processor,
+        ocr_text=text,
+        iou_threshold=iou_threshold,
+    )
+    image = Image.open(io.BytesIO(base64.b64decode(dino_labled_img)))
+    print("finish processing")
+    parsed_content_list_str = "\n".join(parsed_content_list)
+
+    # Encode image to base64
+    buffered = io.BytesIO()
+    image.save(buffered, format="PNG")
+    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
+
+    return ProcessResponse(
+        image=img_str,
+        parsed_content_list=str(parsed_content_list_str),
+        label_coordinates=str(label_coordinates),
+    )
+
+
+@app.post("/process_image", response_model=ProcessResponse)
+async def process_image(
+    image_file: UploadFile = File(...),
+    box_threshold: float = 0.05,
+    iou_threshold: float = 0.1,
+):
+    try:
+        contents = await image_file.read()
+        image_input = Image.open(io.BytesIO(contents)).convert("RGB")
+    except Exception as e:
+        raise HTTPException(status_code=400, detail="Invalid image file")
+
+    response = process(image_input, box_threshold, iou_threshold)
+    return response

download_models.py

@@ -0,0 +1,40 @@
+import os
+from huggingface_hub import hf_hub_download
+import shutil
+
+def download_models():
+    # Create directories if they don't exist
+    os.makedirs("weights/icon_detect", exist_ok=True)
+    os.makedirs("weights/icon_caption_florence", exist_ok=True)
+
+    # Define file mappings (repository path -> local path)
+    files_to_download = {
+        "icon_caption_florence/config.json": "weights/icon_caption_florence/config.json",
+        "icon_caption_florence/generation_config.json": "weights/icon_caption_florence/generation_config.json",
+        "icon_caption_florence/model.safetensors": "weights/icon_caption_florence/model.safetensors",
+        "icon_detect/best.pt": "weights/icon_detect/best.pt"
+    }
+
+    # Download each file
+    for repo_path, local_path in files_to_download.items():
+        if not os.path.exists(local_path):
+            print(f"Downloading {repo_path}...")
+            try:
+                downloaded_file = hf_hub_download(
+                    repo_id="banao-tech/OmniParser",
+                    filename=repo_path,
+                    local_dir="temp"
+                )
+                # Move the file to the correct location
+                os.makedirs(os.path.dirname(local_path), exist_ok=True)
+                shutil.move(downloaded_file, local_path)
+                print(f"Successfully downloaded and moved to {local_path}")
+            except Exception as e:
+                print(f"Error downloading {repo_path}: {str(e)}")
+
+    # Clean up temp directory
+    if os.path.exists("temp"):
+        shutil.rmtree("temp")
+
+if __name__ == "__main__":
+    download_models()

requirements.txt

@@ -0,0 +1,21 @@
+torch==2.0.1  
+easyocr
+torchvision
+supervision==0.18.0
+openai==1.3.5
+transformers
+ultralytics==8.1.24
+azure-identity
+numpy
+opencv-python
+opencv-python-headless
+gradio
+dill
+accelerate
+timm
+einops==0.8.0
+paddlepaddle
+paddleocr
+fastapi
+uvicorn
+huggingface_hub

utils.py

@@ -0,0 +1,535 @@
+# from ultralytics import YOLO
+import os
+import io
+import base64
+import time
+from PIL import Image, ImageDraw, ImageFont
+import json
+import requests
+# utility function
+import os
+from openai import AzureOpenAI
+
+import json
+import sys
+import os
+import cv2
+import numpy as np
+# %matplotlib inline
+from matplotlib import pyplot as plt
+import easyocr
+from paddleocr import PaddleOCR
+reader = easyocr.Reader(['en'])
+paddle_ocr = PaddleOCR(
+    lang='en',  # other lang also available
+    use_angle_cls=False,
+    use_gpu=False,  # using cuda will conflict with pytorch in the same process
+    show_log=False,
+    max_batch_size=1024,
+    use_dilation=True,  # improves accuracy
+    det_db_score_mode='slow',  # improves accuracy
+    rec_batch_num=1024)
+import time
+import base64
+
+import os
+import ast
+import torch
+from typing import Tuple, List
+from torchvision.ops import box_convert
+import re
+from torchvision.transforms import ToPILImage
+import supervision as sv
+import torchvision.transforms as T
+
+
+def get_caption_model_processor(model_name, model_name_or_path="Salesforce/blip2-opt-2.7b", device=None):
+    if not device:
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+    if model_name == "blip2":
+        from transformers import Blip2Processor, Blip2ForConditionalGeneration
+        processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
+        if device == 'cpu':
+            model = Blip2ForConditionalGeneration.from_pretrained(
+            model_name_or_path, device_map=None, torch_dtype=torch.float32
+        ) 
+        else:
+            model = Blip2ForConditionalGeneration.from_pretrained(
+            model_name_or_path, device_map=None, torch_dtype=torch.float16
+        ).to(device)
+    elif model_name == "florence2":
+        from transformers import AutoProcessor, AutoModelForCausalLM 
+        processor = AutoProcessor.from_pretrained("microsoft/Florence-2-base", trust_remote_code=True)
+        if device == 'cpu':
+            model = AutoModelForCausalLM.from_pretrained(model_name_or_path, torch_dtype=torch.float32, trust_remote_code=True)
+        else:
+            model = AutoModelForCausalLM.from_pretrained(model_name_or_path, torch_dtype=torch.float16, trust_remote_code=True).to(device)
+    return {'model': model.to(device), 'processor': processor}
+
+
+def get_yolo_model(model_path):
+    from ultralytics import YOLO
+    # Load the model.
+    model = YOLO(model_path)
+    return model
+
+
+@torch.inference_mode()
+def get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_model_processor, prompt=None, batch_size=32):
+    # Number of samples per batch, --> 256 roughly takes 23 GB of GPU memory for florence model
+
+    to_pil = ToPILImage()
+    if starting_idx:
+        non_ocr_boxes = filtered_boxes[starting_idx:]
+    else:
+        non_ocr_boxes = filtered_boxes
+    croped_pil_image = []
+    for i, coord in enumerate(non_ocr_boxes):
+        xmin, xmax = int(coord[0]*image_source.shape[1]), int(coord[2]*image_source.shape[1])
+        ymin, ymax = int(coord[1]*image_source.shape[0]), int(coord[3]*image_source.shape[0])
+        cropped_image = image_source[ymin:ymax, xmin:xmax, :]
+        croped_pil_image.append(to_pil(cropped_image))
+
+    model, processor = caption_model_processor['model'], caption_model_processor['processor']
+    if not prompt:
+        if 'florence' in model.config.name_or_path:
+            prompt = "<CAPTION>"
+        else:
+            prompt = "The image shows"
+    
+    generated_texts = []
+    device = model.device
+    for i in range(0, len(croped_pil_image), batch_size):
+        start = time.time()
+        batch = croped_pil_image[i:i+batch_size]
+        if model.device.type == 'cuda':
+            inputs = processor(images=batch, text=[prompt]*len(batch), return_tensors="pt").to(device=device, dtype=torch.float16)
+        else:
+            inputs = processor(images=batch, text=[prompt]*len(batch), return_tensors="pt").to(device=device)
+        if 'florence' in model.config.name_or_path:
+            generated_ids = model.generate(input_ids=inputs["input_ids"],pixel_values=inputs["pixel_values"],max_new_tokens=100,num_beams=3, do_sample=False)
+        else:
+            generated_ids = model.generate(**inputs, max_length=100, num_beams=5, no_repeat_ngram_size=2, early_stopping=True, num_return_sequences=1) # temperature=0.01, do_sample=True,
+        generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
+        generated_text = [gen.strip() for gen in generated_text]
+        generated_texts.extend(generated_text)
+    
+    return generated_texts
+
+
+
+def get_parsed_content_icon_phi3v(filtered_boxes, ocr_bbox, image_source, caption_model_processor):
+    to_pil = ToPILImage()
+    if ocr_bbox:
+        non_ocr_boxes = filtered_boxes[len(ocr_bbox):]
+    else:
+        non_ocr_boxes = filtered_boxes
+    croped_pil_image = []
+    for i, coord in enumerate(non_ocr_boxes):
+        xmin, xmax = int(coord[0]*image_source.shape[1]), int(coord[2]*image_source.shape[1])
+        ymin, ymax = int(coord[1]*image_source.shape[0]), int(coord[3]*image_source.shape[0])
+        cropped_image = image_source[ymin:ymax, xmin:xmax, :]
+        croped_pil_image.append(to_pil(cropped_image))
+
+    model, processor = caption_model_processor['model'], caption_model_processor['processor']
+    device = model.device
+    messages = [{"role": "user", "content": "<|image_1|>\ndescribe the icon in one sentence"}] 
+    prompt = processor.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+
+    batch_size = 5  # Number of samples per batch
+    generated_texts = []
+
+    for i in range(0, len(croped_pil_image), batch_size):
+        images = croped_pil_image[i:i+batch_size]
+        image_inputs = [processor.image_processor(x, return_tensors="pt") for x in images]
+        inputs ={'input_ids': [], 'attention_mask': [], 'pixel_values': [], 'image_sizes': []}
+        texts = [prompt] * len(images)
+        for i, txt in enumerate(texts):
+            input = processor._convert_images_texts_to_inputs(image_inputs[i], txt, return_tensors="pt")
+            inputs['input_ids'].append(input['input_ids'])
+            inputs['attention_mask'].append(input['attention_mask'])
+            inputs['pixel_values'].append(input['pixel_values'])
+            inputs['image_sizes'].append(input['image_sizes'])
+        max_len = max([x.shape[1] for x in inputs['input_ids']])
+        for i, v in enumerate(inputs['input_ids']):
+            inputs['input_ids'][i] = torch.cat([processor.tokenizer.pad_token_id * torch.ones(1, max_len - v.shape[1], dtype=torch.long), v], dim=1)
+            inputs['attention_mask'][i] = torch.cat([torch.zeros(1, max_len - v.shape[1], dtype=torch.long), inputs['attention_mask'][i]], dim=1)
+        inputs_cat = {k: torch.concatenate(v).to(device) for k, v in inputs.items()}
+
+        generation_args = { 
+            "max_new_tokens": 25, 
+            "temperature": 0.01, 
+            "do_sample": False, 
+        } 
+        generate_ids = model.generate(**inputs_cat, eos_token_id=processor.tokenizer.eos_token_id, **generation_args) 
+        # # remove input tokens 
+        generate_ids = generate_ids[:, inputs_cat['input_ids'].shape[1]:]
+        response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+        response = [res.strip('\n').strip() for res in response]
+        generated_texts.extend(response)
+
+    return generated_texts
+
+def remove_overlap(boxes, iou_threshold, ocr_bbox=None):
+    assert ocr_bbox is None or isinstance(ocr_bbox, List)
+
+    def box_area(box):
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    def intersection_area(box1, box2):
+        x1 = max(box1[0], box2[0])
+        y1 = max(box1[1], box2[1])
+        x2 = min(box1[2], box2[2])
+        y2 = min(box1[3], box2[3])
+        return max(0, x2 - x1) * max(0, y2 - y1)
+
+    def IoU(box1, box2):
+        intersection = intersection_area(box1, box2)
+        union = box_area(box1) + box_area(box2) - intersection + 1e-6
+        if box_area(box1) > 0 and box_area(box2) > 0:
+            ratio1 = intersection / box_area(box1)
+            ratio2 = intersection / box_area(box2)
+        else:
+            ratio1, ratio2 = 0, 0
+        return max(intersection / union, ratio1, ratio2)
+
+    def is_inside(box1, box2):
+        # return box1[0] >= box2[0] and box1[1] >= box2[1] and box1[2] <= box2[2] and box1[3] <= box2[3]
+        intersection = intersection_area(box1, box2)
+        ratio1 = intersection / box_area(box1)
+        return ratio1 > 0.95
+
+    boxes = boxes.tolist()
+    filtered_boxes = []
+    if ocr_bbox:
+        filtered_boxes.extend(ocr_bbox)
+    # print('ocr_bbox!!!', ocr_bbox)
+    for i, box1 in enumerate(boxes):
+        # if not any(IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2) for j, box2 in enumerate(boxes) if i != j):
+        is_valid_box = True
+        for j, box2 in enumerate(boxes):
+            # keep the smaller box
+            if i != j and IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2):
+                is_valid_box = False
+                break
+        if is_valid_box:
+            # add the following 2 lines to include ocr bbox
+            if ocr_bbox:
+                # only add the box if it does not overlap with any ocr bbox
+                if not any(IoU(box1, box3) > iou_threshold and not is_inside(box1, box3) for k, box3 in enumerate(ocr_bbox)):
+                    filtered_boxes.append(box1)
+            else:
+                filtered_boxes.append(box1)
+    return torch.tensor(filtered_boxes)
+
+
+def remove_overlap_new(boxes, iou_threshold, ocr_bbox=None):
+    '''
+    ocr_bbox format: [{'type': 'text', 'bbox':[x,y], 'interactivity':False, 'content':str }, ...]
+    boxes format: [{'type': 'icon', 'bbox':[x,y], 'interactivity':True, 'content':None }, ...]
+
+    '''
+    assert ocr_bbox is None or isinstance(ocr_bbox, List)
+
+    def box_area(box):
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    def intersection_area(box1, box2):
+        x1 = max(box1[0], box2[0])
+        y1 = max(box1[1], box2[1])
+        x2 = min(box1[2], box2[2])
+        y2 = min(box1[3], box2[3])
+        return max(0, x2 - x1) * max(0, y2 - y1)
+
+    def IoU(box1, box2):
+        intersection = intersection_area(box1, box2)
+        union = box_area(box1) + box_area(box2) - intersection + 1e-6
+        if box_area(box1) > 0 and box_area(box2) > 0:
+            ratio1 = intersection / box_area(box1)
+            ratio2 = intersection / box_area(box2)
+        else:
+            ratio1, ratio2 = 0, 0
+        return max(intersection / union, ratio1, ratio2)
+
+    def is_inside(box1, box2):
+        # return box1[0] >= box2[0] and box1[1] >= box2[1] and box1[2] <= box2[2] and box1[3] <= box2[3]
+        intersection = intersection_area(box1, box2)
+        ratio1 = intersection / box_area(box1)
+        return ratio1 > 0.80
+
+    # boxes = boxes.tolist()
+    filtered_boxes = []
+    if ocr_bbox:
+        filtered_boxes.extend(ocr_bbox)
+    # print('ocr_bbox!!!', ocr_bbox)
+    for i, box1_elem in enumerate(boxes):
+        box1 = box1_elem['bbox']
+        is_valid_box = True
+        for j, box2_elem in enumerate(boxes):
+            # keep the smaller box
+            box2 = box2_elem['bbox']
+            if i != j and IoU(box1, box2) > iou_threshold and box_area(box1) > box_area(box2):
+                is_valid_box = False
+                break
+        if is_valid_box:
+            # add the following 2 lines to include ocr bbox
+            if ocr_bbox:
+                # keep yolo boxes + prioritize ocr label
+                box_added = False
+                for box3_elem in ocr_bbox:
+                    if not box_added:
+                        box3 = box3_elem['bbox']
+                        if is_inside(box3, box1): # ocr inside icon
+                            # box_added = True
+                            # delete the box3_elem from ocr_bbox
+                            try:
+                                filtered_boxes.append({'type': 'text', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': box3_elem['content']})
+                                filtered_boxes.remove(box3_elem)
+                                # print('remove ocr bbox:', box3_elem)
+                            except:
+                                continue
+                            # break
+                        elif is_inside(box1, box3): # icon inside ocr
+                            box_added = True
+                            # try:
+                            #     filtered_boxes.append({'type': 'icon', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': None})
+                            #     filtered_boxes.remove(box3_elem)
+                            # except:
+                            #     continue
+                            break
+                        else:
+                            continue
+                if not box_added:
+                    filtered_boxes.append({'type': 'icon', 'bbox': box1_elem['bbox'], 'interactivity': True, 'content': None})
+                            
+            else:
+                filtered_boxes.append(box1)
+    return filtered_boxes # torch.tensor(filtered_boxes)
+
+
+def load_image(image_path: str) -> Tuple[np.array, torch.Tensor]:
+    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_source = Image.open(image_path).convert("RGB")
+    image = np.asarray(image_source)
+    image_transformed, _ = transform(image_source, None)
+    return image, image_transformed
+
+
+def annotate(image_source: np.ndarray, boxes: torch.Tensor, logits: torch.Tensor, phrases: List[str], text_scale: float, 
+             text_padding=5, text_thickness=2, thickness=3) -> np.ndarray:
+    """    
+    This function annotates an image with bounding boxes and labels.
+
+    Parameters:
+    image_source (np.ndarray): The source image to be annotated.
+    boxes (torch.Tensor): A tensor containing bounding box coordinates. in cxcywh format, pixel scale
+    logits (torch.Tensor): A tensor containing confidence scores for each bounding box.
+    phrases (List[str]): A list of labels for each bounding box.
+    text_scale (float): The scale of the text to be displayed. 0.8 for mobile/web, 0.3 for desktop # 0.4 for mind2web
+
+    Returns:
+    np.ndarray: The annotated image.
+    """
+    h, w, _ = image_source.shape
+    boxes = boxes * torch.Tensor([w, h, w, h])
+    xyxy = box_convert(boxes=boxes, in_fmt="cxcywh", out_fmt="xyxy").numpy()
+    xywh = box_convert(boxes=boxes, in_fmt="cxcywh", out_fmt="xywh").numpy()
+    detections = sv.Detections(xyxy=xyxy)
+
+    labels = [f"{phrase}" for phrase in range(boxes.shape[0])]
+
+    from util.box_annotator import BoxAnnotator 
+    box_annotator = BoxAnnotator(text_scale=text_scale, text_padding=text_padding,text_thickness=text_thickness,thickness=thickness) # 0.8 for mobile/web, 0.3 for desktop # 0.4 for mind2web
+    annotated_frame = image_source.copy()
+    annotated_frame = box_annotator.annotate(scene=annotated_frame, detections=detections, labels=labels, image_size=(w,h))
+
+    label_coordinates = {f"{phrase}": v for phrase, v in zip(phrases, xywh)}
+    return annotated_frame, label_coordinates
+
+
+def predict(model, image, caption, box_threshold, text_threshold):
+    """ Use huggingface model to replace the original model
+    """
+    model, processor = model['model'], model['processor']
+    device = model.device
+
+    inputs = processor(images=image, text=caption, return_tensors="pt").to(device)
+    with torch.no_grad():
+        outputs = model(**inputs)
+
+    results = processor.post_process_grounded_object_detection(
+        outputs,
+        inputs.input_ids,
+        box_threshold=box_threshold, # 0.4,
+        text_threshold=text_threshold, # 0.3,
+        target_sizes=[image.size[::-1]]
+    )[0]
+    boxes, logits, phrases = results["boxes"], results["scores"], results["labels"]
+    return boxes, logits, phrases
+
+
+def predict_yolo(model, image_path, box_threshold, imgsz, scale_img, iou_threshold=0.7):
+    """ Use huggingface model to replace the original model
+    """
+    # model = model['model']
+    if scale_img:
+        result = model.predict(
+        source=image_path,
+        conf=box_threshold,
+        imgsz=imgsz,
+        iou=iou_threshold, # default 0.7
+        )
+    else:
+        result = model.predict(
+        source=image_path,
+        conf=box_threshold,
+        iou=iou_threshold, # default 0.7
+        )
+    boxes = result[0].boxes.xyxy#.tolist() # in pixel space
+    conf = result[0].boxes.conf
+    phrases = [str(i) for i in range(len(boxes))]
+
+    return boxes, conf, phrases
+
+
+def get_som_labeled_img(img_path, model=None, BOX_TRESHOLD = 0.01, output_coord_in_ratio=False, ocr_bbox=None, text_scale=0.4, text_padding=5, draw_bbox_config=None, caption_model_processor=None, ocr_text=[], use_local_semantics=True, iou_threshold=0.9,prompt=None, scale_img=False, imgsz=None, batch_size=None):
+    """ ocr_bbox: list of xyxy format bbox
+    """
+    image_source = Image.open(img_path).convert("RGB")
+    w, h = image_source.size
+    if not imgsz:
+        imgsz = (h, w)
+    # print('image size:', w, h)
+    xyxy, logits, phrases = predict_yolo(model=model, image_path=img_path, box_threshold=BOX_TRESHOLD, imgsz=imgsz, scale_img=scale_img, iou_threshold=0.1)
+    xyxy = xyxy / torch.Tensor([w, h, w, h]).to(xyxy.device)
+    image_source = np.asarray(image_source)
+    phrases = [str(i) for i in range(len(phrases))]
+
+    # annotate the image with labels
+    h, w, _ = image_source.shape
+    if ocr_bbox:
+        ocr_bbox = torch.tensor(ocr_bbox) / torch.Tensor([w, h, w, h])
+        ocr_bbox=ocr_bbox.tolist()
+    else:
+        print('no ocr bbox!!!')
+        ocr_bbox = None
+    # filtered_boxes = remove_overlap(boxes=xyxy, iou_threshold=iou_threshold, ocr_bbox=ocr_bbox)
+    # starting_idx = len(ocr_bbox)
+    # print('len(filtered_boxes):', len(filtered_boxes), starting_idx)
+
+    ocr_bbox_elem = [{'type': 'text', 'bbox':box, 'interactivity':False, 'content':txt} for box, txt in zip(ocr_bbox, ocr_text)]
+    xyxy_elem = [{'type': 'icon', 'bbox':box, 'interactivity':True, 'content':None} for box in xyxy.tolist()]
+    filtered_boxes = remove_overlap_new(boxes=xyxy_elem, iou_threshold=iou_threshold, ocr_bbox=ocr_bbox_elem)
+    
+    # sort the filtered_boxes so that the one with 'content': None is at the end, and get the index of the first 'content': None
+    filtered_boxes_elem = sorted(filtered_boxes, key=lambda x: x['content'] is None)
+    # get the index of the first 'content': None
+    starting_idx = next((i for i, box in enumerate(filtered_boxes_elem) if box['content'] is None), -1)
+    filtered_boxes = torch.tensor([box['bbox'] for box in filtered_boxes_elem])
+
+    
+    # get parsed icon local semantics
+    if use_local_semantics:
+        caption_model = caption_model_processor['model']
+        if 'phi3_v' in caption_model.config.model_type: 
+            parsed_content_icon = get_parsed_content_icon_phi3v(filtered_boxes, ocr_bbox, image_source, caption_model_processor)
+        else:
+            parsed_content_icon = get_parsed_content_icon(filtered_boxes, starting_idx, image_source, caption_model_processor, prompt=prompt,batch_size=batch_size)
+        ocr_text = [f"Text Box ID {i}: {txt}" for i, txt in enumerate(ocr_text)]
+        icon_start = len(ocr_text)
+        parsed_content_icon_ls = []
+        # fill the filtered_boxes_elem None content with parsed_content_icon in order
+        for i, box in enumerate(filtered_boxes_elem):
+            if box['content'] is None:
+                box['content'] = parsed_content_icon.pop(0)
+        for i, txt in enumerate(parsed_content_icon):
+            parsed_content_icon_ls.append(f"Icon Box ID {str(i+icon_start)}: {txt}")
+        parsed_content_merged = ocr_text + parsed_content_icon_ls
+    else:
+        ocr_text = [f"Text Box ID {i}: {txt}" for i, txt in enumerate(ocr_text)]
+        parsed_content_merged = ocr_text
+
+    filtered_boxes = box_convert(boxes=filtered_boxes, in_fmt="xyxy", out_fmt="cxcywh")
+
+    phrases = [i for i in range(len(filtered_boxes))]
+    
+    # draw boxes
+    if draw_bbox_config:
+        annotated_frame, label_coordinates = annotate(image_source=image_source, boxes=filtered_boxes, logits=logits, phrases=phrases, **draw_bbox_config)
+    else:
+        annotated_frame, label_coordinates = annotate(image_source=image_source, boxes=filtered_boxes, logits=logits, phrases=phrases, text_scale=text_scale, text_padding=text_padding)
+    
+    pil_img = Image.fromarray(annotated_frame)
+    buffered = io.BytesIO()
+    pil_img.save(buffered, format="PNG")
+    encoded_image = base64.b64encode(buffered.getvalue()).decode('ascii')
+    if output_coord_in_ratio:
+        # h, w, _ = image_source.shape
+        label_coordinates = {k: [v[0]/w, v[1]/h, v[2]/w, v[3]/h] for k, v in label_coordinates.items()}
+        assert w == annotated_frame.shape[1] and h == annotated_frame.shape[0]
+
+    return encoded_image, label_coordinates, filtered_boxes_elem
+
+
+def get_xywh(input):
+    x, y, w, h = input[0][0], input[0][1], input[2][0] - input[0][0], input[2][1] - input[0][1]
+    x, y, w, h = int(x), int(y), int(w), int(h)
+    return x, y, w, h
+
+def get_xyxy(input):
+    x, y, xp, yp = input[0][0], input[0][1], input[2][0], input[2][1]
+    x, y, xp, yp = int(x), int(y), int(xp), int(yp)
+    return x, y, xp, yp
+
+def get_xywh_yolo(input):
+    x, y, w, h = input[0], input[1], input[2] - input[0], input[3] - input[1]
+    x, y, w, h = int(x), int(y), int(w), int(h)
+    return x, y, w, h
+    
+
+
+def check_ocr_box(image_path, display_img = True, output_bb_format='xywh', goal_filtering=None, easyocr_args=None, use_paddleocr=False):
+    if use_paddleocr:
+        if easyocr_args is None:
+            text_threshold = 0.5
+        else:
+            text_threshold = easyocr_args['text_threshold']
+        result = paddle_ocr.ocr(image_path, cls=False)[0]
+        conf = [item[1] for item in result]
+        coord = [item[0] for item in result if item[1][1] > text_threshold]
+        text = [item[1][0] for item in result if item[1][1] > text_threshold]
+    else:  # EasyOCR
+        if easyocr_args is None:
+            easyocr_args = {}
+        result = reader.readtext(image_path, **easyocr_args)
+        # print('goal filtering pred:', result[-5:])
+        coord = [item[0] for item in result]
+        text = [item[1] for item in result]
+    # read the image using cv2
+    if display_img:
+        opencv_img = cv2.imread(image_path)
+        opencv_img = cv2.cvtColor(opencv_img, cv2.COLOR_RGB2BGR)
+        bb = []
+        for item in coord:
+            x, y, a, b = get_xywh(item)
+            # print(x, y, a, b)
+            bb.append((x, y, a, b))
+            cv2.rectangle(opencv_img, (x, y), (x+a, y+b), (0, 255, 0), 2)
+        
+        # Display the image
+        plt.imshow(opencv_img)
+    else:
+        if output_bb_format == 'xywh':
+            bb = [get_xywh(item) for item in coord]
+        elif output_bb_format == 'xyxy':
+            bb = [get_xyxy(item) for item in coord]
+        # print('bounding box!!!', bb)
+    return (text, bb), goal_filtering
+
+