Rex-Thinker-Demo / tools /inference_tools.py
Mountchicken's picture
Upload 3 files
bf00d99 verified
import re
from typing import Any, Dict, List, Optional, Union
import groundingdino.datasets.transforms as T
import numpy as np
import torch
import torchvision.transforms.functional as F
from groundingdino.util.inference import load_model, predict
from PIL import Image, ImageDraw, ImageFont
from qwen_vl_utils import process_vision_info, smart_resize
class ColorGenerator:
"""A class for generating consistent colors for visualization.
This class provides methods to generate colors either consistently for all elements
or based on text content for better visual distinction.
Args:
color_type (str): Type of color generation strategy. Can be either "same" for consistent color
or "text" for text-based color generation.
"""
def __init__(self, color_type) -> None:
self.color_type = color_type
if color_type == "same":
self.color = tuple((np.random.randint(0, 127, size=3) + 128).tolist())
elif color_type == "text":
np.random.seed(3396)
self.num_colors = 300
self.colors = np.random.randint(0, 127, size=(self.num_colors, 3)) + 128
else:
raise ValueError
def get_color(self, text):
"""Get a color based on the text content or return a consistent color.
Args:
text (str): The text to generate color for.
Returns:
tuple: RGB color values as a tuple.
Raises:
ValueError: If color_type is not supported.
"""
if self.color_type == "same":
return self.color
if self.color_type == "text":
text_hash = hash(text)
index = text_hash % self.num_colors
color = tuple(self.colors[index])
return color
raise ValueError
def visualize(
image_pil: Image,
boxes,
scores,
labels=None,
filter_score=-1,
topN=900,
font_size=15,
draw_width: int = 6,
draw_index: bool = True,
) -> Image:
"""Visualize bounding boxes and labels on an image.
This function draws bounding boxes and their corresponding labels on the input image.
It supports filtering by score, limiting the number of boxes, and customizing the
visualization appearance.
Args:
image_pil (PIL.Image): The input image to draw on.
boxes (List[List[float]]): List of bounding boxes in [x1, y1, x2, y2] format.
scores (List[float]): Confidence scores for each bounding box.
labels (List[str], optional): Labels for each bounding box. Defaults to None.
filter_score (float, optional): Minimum score threshold for visualization. Defaults to -1.
topN (int, optional): Maximum number of boxes to visualize. Defaults to 900.
font_size (int, optional): Font size for labels. Defaults to 15.
draw_width (int, optional): Width of bounding box lines. Defaults to 6.
draw_index (bool, optional): Whether to draw index numbers for unlabeled boxes. Defaults to True.
Returns:
PIL.Image: The image with visualized bounding boxes and labels.
"""
# Get the bounding boxes and labels from the target dictionary
font_path = "tools/Tahoma.ttf"
font = ImageFont.truetype(font_path, font_size)
# Create a PIL ImageDraw object to draw on the input image
draw = ImageDraw.Draw(image_pil)
boxes = boxes[:topN]
scores = scores[:topN]
# Draw boxes and masks for each box and label in the target dictionary
box_idx = 1
color_generaor = ColorGenerator("text")
if labels is None:
labels = [""] * len(boxes)
for box, score, label in zip(boxes, scores, labels):
if score < filter_score:
continue
color = tuple(np.random.randint(0, 255, size=3).tolist())
# Extract the box coordinates
x0, y0, x1, y1 = box
# rescale the box coordinates to the input image size
x0, y0, x1, y1 = int(x0), int(y0), int(x1), int(y1)
if draw_index and label is "":
text = str(box_idx) + f" {label}"
else:
text = str(label)
max_words_per_line = 10
words = text.split()
lines = []
line = ""
for word in words:
if len(line.split()) < max_words_per_line:
line += word + " "
else:
lines.append(line)
line = word + " "
lines.append(line)
text = "\n".join(lines)
draw.rectangle(
[x0, y0, x1, y1], outline=color_generaor.get_color(text), width=draw_width
)
bbox = draw.textbbox((x0, y0), text, font)
box_h = bbox[3] - bbox[1]
box_w = bbox[2] - bbox[0]
y0_text = y0 - box_h - (draw_width * 2)
y1_text = y0 + draw_width
box_idx += 1
if y0_text < 0:
y0_text = 0
y1_text = y0 + 2 * draw_width + box_h
draw.rectangle(
[x0, y0_text, bbox[2] + draw_width * 2, y1_text],
fill=color_generaor.get_color(text),
)
draw.text(
(x0 + draw_width, y0_text),
str(text),
fill="black",
font=font,
)
return image_pil
def compute_iou(box1, box2):
"""Compute Intersection over Union (IoU) between two bounding boxes.
Args:
box1 (List[float]): First bounding box in [x1, y1, x2, y2] format.
box2 (List[float]): Second bounding box in [x1, y1, x2, y2] format.
Returns:
float: IoU score between 0 and 1.
"""
x1 = max(box1[0], box2[0])
y1 = max(box1[1], box2[1])
x2 = min(box1[2], box2[2])
y2 = min(box1[3], box2[3])
inter_area = max(0, x2 - x1) * max(0, y2 - y1)
if inter_area == 0:
return 0.0
box1_area = (box1[2] - box1[0]) * (box1[3] - box1[1])
box2_area = (box2[2] - box2[0]) * (box2[3] - box2[1])
union_area = box1_area + box2_area - inter_area
return inter_area / union_area
def return_maximum_overlap(gt_box, candidate_boxes, min_iou=0.5):
"""Find the best matching box from candidate boxes based on IoU.
Args:
gt_box (List[float]): Ground truth bounding box in [x1, y1, x2, y2] format.
candidate_boxes (List[List[float]]): List of candidate bounding boxes.
min_iou (float, optional): Minimum IoU threshold for matching. Defaults to 0.5.
Returns:
int or None: Index of the best matching box if IoU > min_iou, None otherwise.
"""
max_iou = 0.0
best_box = None
for i, box in enumerate(candidate_boxes):
iou = compute_iou(gt_box, box)
if iou >= min_iou and iou > max_iou:
max_iou = iou
best_box = i
return best_box
def find_best_matched_index(group1, group2):
"""Find the best matching indices between two groups of bounding boxes.
Args:
group1 (List[List[float]]): First group of bounding boxes.
group2 (List[List[float]]): Second group of bounding boxes.
Returns:
List[int]: List of indices (1-based) indicating the best matches from group2 for each box in group1.
"""
labels = []
for box in group1:
best_box = return_maximum_overlap(box, group2)
labels.append(best_box + 1)
return labels
def gdino_load_image(image: Union[str, Image.Image]) -> torch.Tensor:
"""Load and transform image for Grounding DINO model.
Args:
image (Union[str, Image.Image]): Input image path or PIL Image.
Returns:
torch.Tensor: Transformed image tensor ready for model input.
"""
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]),
]
)
if isinstance(image, str):
image_source = Image.open(image).convert("RGB")
else:
image_source = image
image = np.asarray(image_source)
image_transformed, _ = transform(image_source, None)
return image_transformed
def inference_gdino(
image: Image.Image,
prompts: List[str],
gdino_model: Any,
TEXT_TRESHOLD: float = 0.25,
BOX_TRESHOLD: float = 0.25,
) -> torch.Tensor:
"""Process an image with Grounding DINO model to detect objects.
Args:
image (Image.Image): Input PIL image.
prompts (List[str]): List of text prompts for object detection.
gdino_model (Any): The Grounding DINO model instance.
TEXT_TRESHOLD (float, optional): Text confidence threshold. Defaults to 0.25.
BOX_TRESHOLD (float, optional): Box confidence threshold. Defaults to 0.35.
Returns:
List[List[float]]: List of detected bounding boxes in [x1, y1, x2, y2] format.
"""
text_labels = ".".join(prompts)
image_transformed = gdino_load_image(image)
boxes, _, _ = predict(
model=gdino_model,
image=image_transformed,
caption=text_labels,
box_threshold=BOX_TRESHOLD,
text_threshold=TEXT_TRESHOLD,
)
# the output boxes is in the format of (x,y,w,h), in [0,1]
boxes = boxes * torch.tensor([image.width, image.height, image.width, image.height])
# convert to the format of (x1,y1,x2,y2)
boxes = torch.cat(
(boxes[:, :2] - boxes[:, 2:4] / 2, boxes[:, :2] + boxes[:, 2:4] / 2), dim=1
)
return boxes.tolist()
def convert_boxes_from_absolute_to_qwen25_format(gt_boxes, ori_width, ori_height):
"""Convert bounding boxes from absolute coordinates to Qwen-25 format.
This function resizes bounding boxes according to Qwen-25's requirements while
maintaining aspect ratio and pixel constraints.
Args:
gt_boxes (List[List[float]]): List of bounding boxes in absolute coordinates.
ori_width (int): Original image width.
ori_height (int): Original image height.
Returns:
List[List[int]]: Resized bounding boxes in Qwen-25 format.
"""
resized_height, resized_width = smart_resize(
ori_height,
ori_width,
28,
min_pixels=16 * 28 * 28,
max_pixels=1280 * 28 * 28,
)
resized_gt_boxes = []
for box in gt_boxes:
# resize the box
x0, y0, x1, y1 = box
x0 = int(x0 / ori_width * resized_width)
x1 = int(x1 / ori_width * resized_width)
y0 = int(y0 / ori_height * resized_height)
y1 = int(y1 / ori_height * resized_height)
x0 = max(0, min(x0, resized_width - 1))
y0 = max(0, min(y0, resized_height - 1))
x1 = max(0, min(x1, resized_width - 1))
y1 = max(0, min(y1, resized_height - 1))
resized_gt_boxes.append([x0, y0, x1, y1])
return resized_gt_boxes
def parse_json(json_output):
"""Parse JSON string containing coordinate arrays.
Args:
json_output (str): JSON string containing coordinate arrays.
Returns:
List[List[float]]: List of parsed coordinate arrays.
"""
pattern = r"\[([0-9\.]+(?:, ?[0-9\.]+)*)\]"
matches = re.findall(pattern, json_output)
coordinates = [
[float(num) if "." in num else int(num) for num in match.split(",")]
for match in matches
]
return coordinates
def postprocess_and_vis_inference_out(
target_image,
answer,
proposed_box,
gdino_boxes,
font_size,
draw_width,
input_height,
input_width,
):
"""Post-process inference results and create visualization.
This function processes the model output, matches boxes with Grounding DINO results,
and creates visualization images.
Args:
target_image (PIL.Image): Target image for visualization.
answer (str): Model output containing box coordinates.
proposed_box (List[List[float]] or None): Proposed bounding boxes.
gdino_boxes (List[List[float]]): Grounding DINO detected boxes.
font_size (int): Font size for visualization.
draw_width (int): Line width for visualization.
input_height (int): Original input image height.
input_width (int): Original input image width.
Returns:
Tuple[PIL.Image, PIL.Image]: Two visualization images - one for reference boxes
and one for Grounding DINO boxes.
"""
if proposed_box is None:
return target_image, target_image
w, h = target_image.size
json_output = parse_json(answer)
final_boxes = []
input_height = input_height.item()
input_width = input_width.item()
for box in json_output:
x0, y0, x1, y1 = box
x0 = x0 / input_width * w
y0 = y0 / input_height * h
x1 = x1 / input_width * w
y1 = y1 / input_height * h
final_boxes.append([x0, y0, x1, y1])
ref_labels = find_best_matched_index(
final_boxes, gdino_boxes
) # find the best matched index
print("ref_labels", ref_labels)
ref_vis_result = visualize(
target_image.copy(),
final_boxes,
np.ones(len(final_boxes)),
labels=ref_labels,
font_size=font_size,
draw_width=draw_width,
)
dinox_vis_result = visualize(
target_image.copy(),
gdino_boxes,
np.ones(len(gdino_boxes)),
font_size=font_size,
draw_width=draw_width,
)
return ref_vis_result, dinox_vis_result