gpcv_incontext_bench / visualize_results.py
insomnia7's picture
Add files using upload-large-folder tool
2c986a2 verified
Raw
History Blame Contribute Delete
4.31 kB
import json
import re
import argparse
from pathlib import Path
from PIL import Image, ImageDraw
GRID_SIZE = 1000
def load_json(path):
with open(path, "r") as f:
return json.load(f)
def extract_ref_boxes(prompt):
matches = re.findall(r"<x(\d+)><y(\d+)><x(\d+)><y(\d+)>", prompt)
return [[int(x1), int(y1), int(x2), int(y2)] for x1, y1, x2, y2 in matches]
def grid_to_pixel(boxes, img_w, img_h):
"""将 1000-grid 坐标转为像素坐标"""
return [[
int(x1 * img_w / GRID_SIZE),
int(y1 * img_h / GRID_SIZE),
int(x2 * img_w / GRID_SIZE),
int(y2 * img_h / GRID_SIZE),
] for x1, y1, x2, y2 in boxes]
def draw_scaled_boxes(draw, boxes, off_x, off_y, scale, color, width=3):
for x1, y1, x2, y2 in boxes:
draw.rectangle([
off_x + x1 * scale, off_y + y1 * scale,
off_x + x2 * scale, off_y + y2 * scale,
], outline=color, width=width)
def visualize_single(data_path, output_path, max_images=None, cols=4):
data = load_json(data_path)
if max_images:
data = data[:max_images]
n = len(data)
rows = (n + cols - 1) // cols
images = []
for item in data:
img = Image.open(item["image"]).convert("RGB")
images.append(img)
max_w = max(im.width for im in images)
max_h = max(im.height for im in images)
header_h = 60
cell_pad = 8
canvas_w = cols * (max_w + cell_pad) + cell_pad
canvas_h = rows * (max_h + cell_pad) + cell_pad + header_h
canvas = Image.new("RGB", (canvas_w, canvas_h), (240, 240, 240))
draw = ImageDraw.Draw(canvas)
legend_x = cell_pad
legend_y = cell_pad
draw.rectangle([legend_x, legend_y, legend_x + 20, legend_y + 12], fill=None, outline=(0, 200, 0), width=3)
draw.text((legend_x + 26, legend_y - 1), "GT", fill=(0, 0, 0))
draw.rectangle([legend_x + 80, legend_y, legend_x + 100, legend_y + 12], fill=None, outline=(220, 20, 20), width=3)
draw.text((legend_x + 106, legend_y - 1), "Pred", fill=(0, 0, 0))
draw.rectangle([legend_x + 160, legend_y, legend_x + 180, legend_y + 12], fill=None, outline=(0, 120, 220), width=3)
draw.text((legend_x + 186, legend_y - 1), "Ref", fill=(0, 0, 0))
draw.text((legend_x + 240, legend_y - 1), f"Total: {n} images", fill=(0, 0, 0))
for idx, (item, img) in enumerate(zip(data, images)):
r = idx // cols
c = idx % cols
ox = cell_pad + c * (max_w + cell_pad)
oy = header_h + cell_pad + r * (max_h + cell_pad)
scale = min(max_w / img.width, max_h / img.height)
new_w = int(img.width * scale)
new_h = int(img.height * scale)
img_resized = img.resize((new_w, new_h))
off_x = ox + (max_w - new_w) // 2
off_y = oy + (max_h - new_h) // 2
canvas.paste(img_resized, (off_x, off_y))
gt_boxes = grid_to_pixel(item.get("gt_bboxes", []), img.width, img.height)
pred_boxes = grid_to_pixel(item.get("pred_bboxes", []), img.width, img.height)
ref_boxes = grid_to_pixel(extract_ref_boxes(item.get("prompt", "")), img.width, img.height)
draw_scaled_boxes(draw, ref_boxes, off_x, off_y, scale, (0, 120, 220), width=2)
draw_scaled_boxes(draw, gt_boxes, off_x, off_y, scale, (0, 200, 0), width=3)
draw_scaled_boxes(draw, pred_boxes, off_x, off_y, scale, (220, 20, 20), width=3)
fname = Path(item["image"]).name
stats = f"GT:{len(gt_boxes)} Pred:{len(pred_boxes)}"
draw.text((ox + 4, oy + max_h - 18), f"{fname} {stats}", fill=(0, 0, 0))
canvas.save(output_path)
print(f"保存到: {output_path}")
def main():
parser = argparse.ArgumentParser(description="可视化 GT 和 Pred bboxes")
parser.add_argument("--input", "-i", required=True, help="简化结果 JSON 路径")
parser.add_argument("--output", "-o", default="contact_sheet.jpg", help="输出图片路径")
parser.add_argument("--max", "-n", type=int, default=0, help="最多可视化几张 (0=全部)")
parser.add_argument("--cols", type=int, default=4, help="每行列数")
args = parser.parse_args()
visualize_single(
args.input,
args.output,
max_images=args.max if args.max > 0 else None,
cols=args.cols,
)
if __name__ == "__main__":
main()