layout / compare /data /compare.py
hassanshka's picture
Add test_combined_models.py and compare/ folder (excluding cvat_project_7_export and Annika 2 folders)
0a216c0
"""
Main comparison script: Compare old models vs new models vs ground truth.
Calculates mAP@50, mAP@[.50:.95], Precision, Recall.
Creates side-by-side visualization.
"""
import os
import json
import sys
from pathlib import Path
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt
import matplotlib.patches as patches
import matplotlib.colors as mcolors
try:
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
HAS_PYCOCOTOOLS = True
except ImportError:
HAS_PYCOCOTOOLS = False
print("Warning: pycocotools not available. Metrics calculation will be limited.")
COCO = None
COCOeval = None
import tempfile
# Add project root to path
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
PROJECT_ROOT = os.path.dirname(os.path.dirname(SCRIPT_DIR))
sys.path.insert(0, PROJECT_ROOT)
sys.path.insert(0, SCRIPT_DIR)
from original_annotations import load_ground_truth
from old_models import process_dataset as process_old_models
from new_models import process_dataset as process_new_models
def draw_coco_annotations_simple(image_path, coco_json, title="", ax=None):
"""
Draw COCO annotations on image (simpler version for comparison).
"""
if ax is None:
fig, ax = plt.subplots(1, 1, figsize=(10, 14))
img = Image.open(image_path).convert("RGB")
ax.imshow(img)
ax.set_title(title, fontsize=14, fontweight='bold')
ax.axis("off")
if not coco_json.get("images"):
return ax
img_info = coco_json["images"][0]
img_id = img_info["id"]
anns = [a for a in coco_json["annotations"] if a["image_id"] == img_id]
id_to_name = {c["id"]: c["name"] for c in coco_json["categories"]}
# Color map
colors = plt.cm.tab20(np.linspace(0, 1, 20))
color_map = {}
# Track label positions to avoid overlap
placed_labels = []
def find_label_position(bbox, text_width, text_height, image_width, image_height):
"""Find a good position for label to avoid overlap."""
x, y, w, h = bbox
candidates = [
(x, y - text_height - 5), # Above top-left
(x, y), # Top-left corner
(x + w - text_width, y), # Top-right corner
(x, y + h + 5), # Below bottom-left
]
for pos_x, pos_y in candidates:
# Check if position is within image bounds
if pos_x < 0 or pos_y < 0 or pos_x + text_width > image_width or pos_y + text_height > image_height:
continue
# Check overlap with existing labels
overlap = False
for placed_x, placed_y, placed_w, placed_h in placed_labels:
if not (pos_x + text_width < placed_x or pos_x > placed_x + placed_w or
pos_y + text_height < placed_y or pos_y > placed_y + placed_h):
overlap = True
break
if not overlap:
return pos_x, pos_y
# If all positions overlap, use top-left anyway
return x, y
img_width, img_height = img.size
for ann in anns:
name = id_to_name.get(ann["category_id"], f"cls_{ann['category_id']}")
# Get or assign color
if name not in color_map:
color_idx = len(color_map) % len(colors)
color_map[name] = colors[color_idx]
color = color_map[name]
# Get bbox for label positioning
bbox = ann.get("bbox", [0, 0, 0, 0])
if not bbox or len(bbox) < 4:
# Try to get bbox from segmentation
segs = ann.get("segmentation", [])
if segs and isinstance(segs, list) and len(segs) > 0:
if isinstance(segs[0], list) and len(segs[0]) >= 6:
coords = segs[0]
xs = coords[0::2]
ys = coords[1::2]
bbox = [min(xs), min(ys), max(xs) - min(xs), max(ys) - min(ys)]
else:
continue
else:
continue
x, y, w, h = bbox
# Draw segmentation or bbox
segs = ann.get("segmentation", [])
if segs and isinstance(segs, list) and len(segs) > 0:
if isinstance(segs[0], list) and len(segs[0]) >= 6:
# Polygon
coords = segs[0]
xs = coords[0::2]
ys = coords[1::2]
poly = patches.Polygon(
list(zip(xs, ys)),
closed=True,
edgecolor=color,
facecolor=color,
linewidth=2,
alpha=0.3,
)
ax.add_patch(poly)
# Edge
poly_edge = patches.Polygon(
list(zip(xs, ys)),
closed=True,
edgecolor=color,
facecolor="none",
linewidth=2,
alpha=0.8,
)
ax.add_patch(poly_edge)
else:
# Bbox
rect = patches.Rectangle(
(x, y),
w,
h,
edgecolor=color,
facecolor=color,
linewidth=2,
alpha=0.3,
)
ax.add_patch(rect)
# Edge
rect_edge = patches.Rectangle(
(x, y),
w,
h,
edgecolor=color,
facecolor="none",
linewidth=2,
alpha=0.8,
)
ax.add_patch(rect_edge)
# Add label
# Estimate text size (approximate)
text_width = len(name) * 7 # Approximate character width
text_height = 12 # Approximate text height
label_x, label_y = find_label_position(bbox, text_width, text_height, img_width, img_height)
placed_labels.append((label_x, label_y, text_width, text_height))
# Draw label with background
# Convert color to RGB tuple if it's an array
if isinstance(color, np.ndarray):
edge_color = tuple(color[:3])
elif isinstance(color, (list, tuple)) and len(color) >= 3:
edge_color = tuple(color[:3])
else:
edge_color = color
ax.text(
label_x,
label_y,
name,
color='black',
fontsize=9,
fontweight='bold',
bbox=dict(
boxstyle="round,pad=0.3",
facecolor="white",
edgecolor=edge_color,
linewidth=2,
alpha=0.9,
),
zorder=10, # Ensure labels are on top
)
return ax
def validate_and_fix_annotation(ann, img_width, img_height):
"""
Validate and fix annotation segmentation/bbox.
Converts bbox to polygon if segmentation is missing or invalid.
"""
segs = ann.get("segmentation", [])
bbox = ann.get("bbox", [0, 0, 0, 0])
# Check if segmentation is valid
has_valid_seg = False
if segs and isinstance(segs, list) and len(segs) > 0:
# Check if it's a polygon (list of coordinates)
if isinstance(segs[0], list) and len(segs[0]) >= 6:
# Valid polygon
has_valid_seg = True
# Check if it's RLE (dict)
elif isinstance(segs, dict) or (isinstance(segs, list) and len(segs) > 0 and isinstance(segs[0], dict)):
# RLE format - assume valid
has_valid_seg = True
# If no valid segmentation, create polygon from bbox
if not has_valid_seg and len(bbox) == 4 and bbox[2] > 0 and bbox[3] > 0:
x, y, w, h = bbox
# Create polygon from bbox: [x, y, x+w, y, x+w, y+h, x, y+h]
polygon = [x, y, x + w, y, x + w, y + h, x, y + h]
ann["segmentation"] = [polygon]
# Update area if needed
if ann.get("area", 0) == 0:
ann["area"] = w * h
has_valid_seg = True
return has_valid_seg
def filter_valid_annotations(coco_dict):
"""
Filter out annotations with invalid segmentation/bbox.
Convert bbox-only annotations to polygon format.
"""
# Get image dimensions
img_id_to_size = {}
for img in coco_dict["images"]:
img_id_to_size[img["id"]] = (img["width"], img["height"])
valid_annotations = []
for ann in coco_dict["annotations"]:
img_id = ann["image_id"]
if img_id in img_id_to_size:
img_width, img_height = img_id_to_size[img_id]
if validate_and_fix_annotation(ann, img_width, img_height):
valid_annotations.append(ann)
coco_dict["annotations"] = valid_annotations
return coco_dict
def calculate_metrics(gt_coco, pred_coco, output_dir):
"""
Calculate mAP@50, mAP@[.50:.95], Precision, Recall using pycocotools.
Args:
gt_coco: Ground truth COCO format dict
pred_coco: Predictions COCO format dict
output_dir: Directory to save results
Returns:
Dictionary with metrics
"""
if not HAS_PYCOCOTOOLS:
return {
'mAP@50': 0.0,
'mAP@[.50:.95]': 0.0,
'Precision': 0.0,
'Recall': 0.0,
'F1': 0.0,
'error': 'pycocotools not available'
}
# Filter and fix invalid annotations
gt_coco_clean = filter_valid_annotations(gt_coco.copy())
pred_coco_clean = filter_valid_annotations(pred_coco.copy())
if len(gt_coco_clean["annotations"]) == 0:
print("Warning: No valid ground truth annotations after filtering")
return {
'mAP@50': 0.0,
'mAP@[.50:.95]': 0.0,
'Precision': 0.0,
'Recall': 0.0,
'F1': 0.0,
'error': 'No valid GT annotations'
}
if len(pred_coco_clean["annotations"]) == 0:
print("Warning: No valid prediction annotations after filtering")
return {
'mAP@50': 0.0,
'mAP@[.50:.95]': 0.0,
'Precision': 0.0,
'Recall': 0.0,
'F1': 0.0,
'error': 'No valid prediction annotations'
}
# Save to temporary JSON files for pycocotools
gt_file = os.path.join(output_dir, "gt_temp.json")
pred_file = os.path.join(output_dir, "pred_temp.json")
with open(gt_file, 'w') as f:
json.dump(gt_coco_clean, f)
with open(pred_file, 'w') as f:
json.dump(pred_coco_clean, f)
# Load with pycocotools
try:
gt_coco_obj = COCO(gt_file)
pred_coco_obj = COCO(pred_file)
except Exception as e:
print(f"Error loading COCO files: {e}")
return {
'mAP@50': 0.0,
'mAP@[.50:.95]': 0.0,
'Precision': 0.0,
'Recall': 0.0,
'F1': 0.0,
'error': f'COCO load error: {str(e)}'
}
# Get all image IDs
img_ids = sorted(gt_coco_obj.getImgIds())
if len(img_ids) == 0:
return {
'mAP@50': 0.0,
'mAP@[.50:.95]': 0.0,
'Precision': 0.0,
'Recall': 0.0,
'F1': 0.0,
'error': 'No images in GT'
}
# Get all category IDs from ground truth
cat_ids = sorted(gt_coco_obj.getCatIds())
# Try segmentation evaluation first, fall back to bbox if it fails
eval_type = 'segm'
try:
coco_eval = COCOeval(gt_coco_obj, pred_coco_obj, eval_type)
coco_eval.params.imgIds = img_ids
coco_eval.params.catIds = cat_ids
coco_eval.evaluate()
coco_eval.accumulate()
coco_eval.summarize()
# Extract metrics
metrics = {
'mAP@50': float(coco_eval.stats[1]), # mAP@0.50
'mAP@[.50:.95]': float(coco_eval.stats[0]), # mAP@[.50:.95]
'mAP@75': float(coco_eval.stats[2]), # mAP@0.75
'mAP_small': float(coco_eval.stats[3]),
'mAP_medium': float(coco_eval.stats[4]),
'mAP_large': float(coco_eval.stats[5]),
'mAR_1': float(coco_eval.stats[6]),
'mAR_10': float(coco_eval.stats[7]),
'mAR_100': float(coco_eval.stats[8]),
'mAR_small': float(coco_eval.stats[9]),
'mAR_medium': float(coco_eval.stats[10]),
'mAR_large': float(coco_eval.stats[11]),
}
# Calculate Precision and Recall
precision = metrics['mAP@50'] # Approximate
recall = metrics['mAR_100'] # Maximum recall with 100 detections
metrics['Precision'] = precision
metrics['Recall'] = recall
metrics['F1'] = 2 * (precision * recall) / (precision + recall) if (precision + recall) > 0 else 0
except Exception as e:
print(f"Error during {eval_type} evaluation: {e}")
# Try bbox evaluation as fallback
try:
print("Trying bbox evaluation as fallback...")
coco_eval = COCOeval(gt_coco_obj, pred_coco_obj, 'bbox')
coco_eval.params.imgIds = img_ids
coco_eval.params.catIds = cat_ids
coco_eval.evaluate()
coco_eval.accumulate()
coco_eval.summarize()
metrics = {
'mAP@50': float(coco_eval.stats[1]),
'mAP@[.50:.95]': float(coco_eval.stats[0]),
'mAP@75': float(coco_eval.stats[2]),
'mAP_small': float(coco_eval.stats[3]),
'mAP_medium': float(coco_eval.stats[4]),
'mAP_large': float(coco_eval.stats[5]),
'mAR_1': float(coco_eval.stats[6]),
'mAR_10': float(coco_eval.stats[7]),
'mAR_100': float(coco_eval.stats[8]),
'mAR_small': float(coco_eval.stats[9]),
'mAR_medium': float(coco_eval.stats[10]),
'mAR_large': float(coco_eval.stats[11]),
}
precision = metrics['mAP@50']
recall = metrics['mAR_100']
metrics['Precision'] = precision
metrics['Recall'] = recall
metrics['F1'] = 2 * (precision * recall) / (precision + recall) if (precision + recall) > 0 else 0
metrics['eval_type'] = 'bbox' # Note that we used bbox evaluation
except Exception as e2:
print(f"Error during bbox evaluation: {e2}")
import traceback
traceback.print_exc()
metrics = {
'mAP@50': 0.0,
'mAP@[.50:.95]': 0.0,
'Precision': 0.0,
'Recall': 0.0,
'F1': 0.0,
'error': f'{eval_type} error: {str(e)}, bbox error: {str(e2)}'
}
return metrics
def create_comparison_visualization(image_path, gt_coco, old_coco, new_coco, output_path):
"""
Create side-by-side comparison: Original + GT | Old Models | New Models
"""
fig, axes = plt.subplots(1, 3, figsize=(30, 10))
# Left: Original image with ground truth
draw_coco_annotations_simple(image_path, gt_coco, "Ground Truth", axes[0])
# Middle: Old models
draw_coco_annotations_simple(image_path, old_coco, "Old Models", axes[1])
# Right: New models
draw_coco_annotations_simple(image_path, new_coco, "New Models", axes[2])
plt.tight_layout()
plt.savefig(output_path, dpi=150, bbox_inches='tight')
plt.close()
print(f"Saved comparison visualization to {output_path}")
def align_categories(gt_coco, pred_coco):
"""
Align category IDs between GT and predictions.
Maps prediction categories to GT categories by name.
"""
# Create name to ID maps
gt_name_to_id = {c["name"]: c["id"] for c in gt_coco["categories"]}
pred_name_to_id = {c["name"]: c["id"] for c in pred_coco["categories"]}
# Create mapping from pred category ID to GT category ID
pred_to_gt_map = {}
for pred_name, pred_id in pred_name_to_id.items():
if pred_name in gt_name_to_id:
pred_to_gt_map[pred_id] = gt_name_to_id[pred_name]
else:
# If category doesn't exist in GT, skip it
print(f"Warning: Category '{pred_name}' not in ground truth, skipping...")
# Update prediction annotations
new_anns = []
for ann in pred_coco["annotations"]:
old_cat_id = ann["category_id"]
if old_cat_id in pred_to_gt_map:
new_ann = ann.copy()
new_ann["category_id"] = pred_to_gt_map[old_cat_id]
new_anns.append(new_ann)
pred_coco["annotations"] = new_anns
# Update categories to match GT
pred_coco["categories"] = [
c for c in gt_coco["categories"]
if c["name"] in pred_name_to_id
]
return pred_coco
def main():
"""
Main comparison function.
"""
# Paths
data_dir = os.path.join(SCRIPT_DIR, "Aleyna 1 (2024)")
xml_path = os.path.join(data_dir, "Annotations", "annotations.xml")
images_dir = os.path.join(data_dir, "Images")
output_dir = os.path.join(SCRIPT_DIR, "results")
os.makedirs(output_dir, exist_ok=True)
print("=" * 60)
print("COMPARISON: Old Models vs New Models vs Ground Truth")
print("=" * 60)
# 1. Load ground truth
print("\n[1/4] Loading ground truth annotations...")
gt_coco = load_ground_truth(xml_path, images_dir)
print(f" βœ“ Loaded {len(gt_coco['images'])} images")
print(f" βœ“ Loaded {len(gt_coco['annotations'])} annotations")
print(f" βœ“ Categories: {[c['name'] for c in gt_coco['categories']]}")
# Save GT
gt_output = os.path.join(output_dir, "ground_truth.json")
with open(gt_output, 'w') as f:
json.dump(gt_coco, f, indent=2)
print(f" βœ“ Saved to {gt_output}")
# 2. Run old models
print("\n[2/4] Running old models...")
old_output_dir = os.path.join(output_dir, "old_models")
os.makedirs(old_output_dir, exist_ok=True)
old_coco = process_old_models(images_dir, old_output_dir)
print(f" βœ“ Processed {len(old_coco['images'])} images")
print(f" βœ“ Generated {len(old_coco['annotations'])} annotations")
old_output = os.path.join(output_dir, "old_models_merged.json")
with open(old_output, 'w') as f:
json.dump(old_coco, f, indent=2)
print(f" βœ“ Saved to {old_output}")
# 3. Run new models
print("\n[3/4] Running new models...")
new_output_dir = os.path.join(output_dir, "new_models")
os.makedirs(new_output_dir, exist_ok=True)
new_coco = process_new_models(images_dir, new_output_dir)
print(f" βœ“ Processed {len(new_coco['images'])} images")
print(f" βœ“ Generated {len(new_coco['annotations'])} annotations")
new_output = os.path.join(output_dir, "new_models_merged.json")
with open(new_output, 'w') as f:
json.dump(new_coco, f, indent=2)
print(f" βœ“ Saved to {new_output}")
# 4. Calculate metrics
print("\n[4/4] Calculating metrics...")
# Align categories
old_coco_aligned = align_categories(gt_coco.copy(), old_coco.copy())
new_coco_aligned = align_categories(gt_coco.copy(), new_coco.copy())
# Calculate metrics for old models
print("\n Calculating metrics for OLD MODELS...")
old_metrics = calculate_metrics(gt_coco, old_coco_aligned, output_dir)
print(f" mAP@50: {old_metrics['mAP@50']:.4f}")
print(f" mAP@[.50:.95]: {old_metrics['mAP@[.50:.95]']:.4f}")
print(f" Precision: {old_metrics['Precision']:.4f}")
print(f" Recall: {old_metrics['Recall']:.4f}")
# Calculate metrics for new models
print("\n Calculating metrics for NEW MODELS...")
new_metrics = calculate_metrics(gt_coco, new_coco_aligned, output_dir)
print(f" mAP@50: {new_metrics['mAP@50']:.4f}")
print(f" mAP@[.50:.95]: {new_metrics['mAP@[.50:.95]']:.4f}")
print(f" Precision: {new_metrics['Precision']:.4f}")
print(f" Recall: {new_metrics['Recall']:.4f}")
# Save metrics
metrics_output = os.path.join(output_dir, "metrics.json")
with open(metrics_output, 'w') as f:
json.dump({
'old_models': old_metrics,
'new_models': new_metrics
}, f, indent=2)
print(f"\n βœ“ Saved metrics to {metrics_output}")
# 5. Create visualizations for each image
print("\n[5/5] Creating comparison visualizations...")
vis_dir = os.path.join(output_dir, "visualizations")
os.makedirs(vis_dir, exist_ok=True)
for img_info in gt_coco["images"]:
image_name = img_info["file_name"]
image_path = os.path.join(images_dir, image_name)
if not os.path.exists(image_path):
continue
# Get COCO for this image
img_id = img_info["id"]
# Filter annotations for this image
gt_img_coco = {
"images": [img_info],
"annotations": [a for a in gt_coco["annotations"] if a["image_id"] == img_id],
"categories": gt_coco["categories"]
}
old_img_coco = {
"images": [img_info],
"annotations": [a for a in old_coco["annotations"] if a["image_id"] == img_id],
"categories": old_coco["categories"]
}
new_img_coco = {
"images": [img_info],
"annotations": [a for a in new_coco["annotations"] if a["image_id"] == img_id],
"categories": new_coco["categories"]
}
# Create visualization
output_path = os.path.join(vis_dir, f"{Path(image_name).stem}_comparison.png")
create_comparison_visualization(
image_path,
gt_img_coco,
old_img_coco,
new_img_coco,
output_path
)
print(f"\n βœ“ Saved visualizations to {vis_dir}")
print("\n" + "=" * 60)
print("COMPARISON COMPLETE!")
print("=" * 60)
print(f"\nResults saved to: {output_dir}")
if __name__ == "__main__":
main()