RFDETR-Soccernet / inference.py
julianzu9612's picture
Add RF-DETR SoccerNet model (85.7% mAP@50) with DataFrame API
7e56761 verified
#!/usr/bin/env python3
"""
RF-DETR SoccerNet Inference - Professional Hugging Face Integration
A state-of-the-art RF-DETR-Large model fine-tuned on the SoccerNet-Tracking dataset
for detecting objects in soccer videos. Returns detections as pandas DataFrame.
Classes: ball, player, referee, goalkeeper
Performance: 85.7% mAP@50, 49.8% mAP
"""
import cv2
import pandas as pd
import numpy as np
import torch
from rfdetr import RFDETRBase
from PIL import Image
from typing import Union, Optional, List, Dict, Tuple
import os
from tqdm import tqdm
import time
import json
from pathlib import Path
import warnings
# Suppress warnings for cleaner output
warnings.filterwarnings("ignore")
class RFDETRSoccerNet:
"""
RF-DETR model trained on SoccerNet dataset for soccer video analysis.
Returns detections as pandas DataFrame with comprehensive metadata.
Performance:
- mAP@50: 85.7%
- mAP: 49.8%
- Classes: ball, player, referee, goalkeeper
- Training: 42,750 images, NVIDIA A100 40GB, ~14 hours
"""
def __init__(self, model_path: str = "weights/checkpoint_best_regular.pth", device: str = "auto"):
"""
Initialize the RF-DETR SoccerNet model.
Args:
model_path: Path to the model checkpoint (default: "weights/checkpoint_best_regular.pth")
device: Device to use ("cuda", "cpu", or "auto" for automatic selection)
"""
# Determine device
if device == "auto":
self.device = "cuda" if torch.cuda.is_available() else "cpu"
else:
self.device = device
print(f"πŸš€ Initializing RF-DETR SoccerNet on {self.device.upper()}")
# Class configuration for SoccerNet dataset
self.class_names = ['ball', 'player', 'referee', 'goalkeeper']
self.num_classes = len(self.class_names)
# Model metadata
self.model_info = {
"architecture": "RF-DETR-Large",
"parameters": "128M",
"input_size": [1280, 1280],
"performance": {
"mAP@50": 0.857,
"mAP": 0.498,
"mAP@75": 0.520
}
}
# Load model
self.model_path = Path(model_path)
self.model = None
self._load_model()
print("βœ… RF-DETR SoccerNet ready for inference!")
def _load_model(self):
"""Load the RF-DETR model with trained checkpoint."""
try:
print(f"πŸ“¦ Loading model from {self.model_path}...")
# Initialize base model
self.model = RFDETRBase()
# Reinitialize detection head for 4 classes (critical for compatibility)
print(f"πŸ”§ Reinitializing detection head for {self.num_classes} classes...")
self.model.model.model.reinitialize_detection_head(self.num_classes)
# Load checkpoint
if self.model_path.exists():
checkpoint = torch.load(str(self.model_path), map_location=self.device, weights_only=False)
# Extract model state
if 'model' in checkpoint:
model_state = checkpoint['model']
elif 'model_state_dict' in checkpoint:
model_state = checkpoint['model_state_dict']
else:
model_state = checkpoint
# Load state dict
self.model.model.model.load_state_dict(model_state)
# Show checkpoint info
if 'best_mAP' in checkpoint:
print(f"πŸ“Š Model mAP: {checkpoint['best_mAP']:.3f}")
if 'epoch' in checkpoint:
print(f"πŸ”„ Trained epochs: {checkpoint['epoch']}")
else:
raise FileNotFoundError(f"Checkpoint not found: {self.model_path}")
# Move to device and set eval mode
self.model.model.model.to(self.device)
self.model.model.model.eval()
print(f"βœ… Model loaded successfully!")
except Exception as e:
print(f"❌ Error loading model: {e}")
raise
def process_video(self,
video_path: str,
confidence_threshold: float = 0.5,
frame_skip: int = 1,
max_frames: Optional[int] = None,
save_results: bool = False,
output_dir: Optional[str] = None) -> pd.DataFrame:
"""
Process a video and return detections as DataFrame.
Args:
video_path: Path to input video
confidence_threshold: Minimum confidence for detections (0.0-1.0)
frame_skip: Process every N frames (1 = all frames)
max_frames: Maximum frames to process (None = all)
save_results: Whether to save results to file
output_dir: Directory to save results (optional)
Returns:
DataFrame with columns: frame, timestamp, class_name, x1, y1, x2, y2, width, height, confidence
"""
print(f"🎬 Processing video: {video_path}")
# Open video
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
raise ValueError(f"Could not open video: {video_path}")
# Video metadata
total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
fps = cap.get(cv2.CAP_PROP_FPS)
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
print(f"πŸ“Ή Video info: {total_frames} frames, {fps:.2f} FPS, {width}x{height}")
# Process frames
results = []
frame_count = 0
processed_count = 0
start_time = time.time()
frames_to_process = min(total_frames, max_frames) if max_frames else total_frames
pbar = tqdm(total=frames_to_process, desc="Processing frames", unit="frame")
while cap.isOpened() and frame_count < frames_to_process:
ret, frame = cap.read()
if not ret:
break
# Process frame based on skip rate
if frame_count % frame_skip == 0:
# Convert to RGB for model
frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
pil_image = Image.fromarray(frame_rgb)
# Run inference
with torch.no_grad():
detections = self.model.predict(pil_image, threshold=confidence_threshold)
# Process detections
if detections is not None and len(detections) > 0:
for i in range(len(detections)):
try:
class_id = int(detections.class_id[i])
if 0 <= class_id < len(self.class_names):
x1, y1, x2, y2 = detections.xyxy[i].tolist()
results.append({
'frame': frame_count,
'timestamp': frame_count / fps,
'class_name': self.class_names[class_id],
'class_id': class_id,
'x1': float(x1),
'y1': float(y1),
'x2': float(x2),
'y2': float(y2),
'width': float(x2 - x1),
'height': float(y2 - y1),
'confidence': float(detections.confidence[i]),
'center_x': float((x1 + x2) / 2),
'center_y': float((y1 + y2) / 2),
'area': float((x2 - x1) * (y2 - y1))
})
except Exception as e:
print(f"⚠️ Error processing detection {i}: {e}")
continue
processed_count += 1
frame_count += 1
pbar.update(1)
cap.release()
pbar.close()
# Create DataFrame
df = pd.DataFrame(results)
# Processing summary
processing_time = time.time() - start_time
fps_processed = processed_count / processing_time if processing_time > 0 else 0
print(f"\nβœ… Processing complete!")
print(f"πŸ“ˆ Stats:")
print(f" - Total frames: {frame_count:,}")
print(f" - Frames processed: {processed_count:,}")
print(f" - Processing time: {processing_time:.1f}s")
print(f" - Processing speed: {fps_processed:.1f} FPS")
print(f" - Total detections: {len(df):,}")
if len(df) > 0:
print(f"\n🎯 Detections by class:")
class_counts = df['class_name'].value_counts()
for class_name, count in class_counts.items():
percentage = (count / len(df)) * 100
print(f" - {class_name}: {count:,} ({percentage:.1f}%)")
# Save results if requested
if save_results:
self._save_video_results(df, video_path, output_dir, {
'total_frames': frame_count,
'processed_frames': processed_count,
'processing_time': processing_time,
'fps_processed': fps_processed,
'video_fps': fps,
'video_resolution': f"{width}x{height}"
})
return df
def process_image(self,
image_path: str,
confidence_threshold: float = 0.5) -> pd.DataFrame:
"""
Process a single image and return detections as DataFrame.
Args:
image_path: Path to input image
confidence_threshold: Minimum confidence for detections
Returns:
DataFrame with columns: class_name, x1, y1, x2, y2, width, height, confidence
"""
print(f"πŸ–ΌοΈ Processing image: {image_path}")
# Load and validate image
if not os.path.exists(image_path):
raise FileNotFoundError(f"Image not found: {image_path}")
image = cv2.imread(image_path)
if image is None:
raise ValueError(f"Could not load image: {image_path}")
# Convert to RGB
image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
pil_image = Image.fromarray(image_rgb)
# Run inference
start_time = time.time()
with torch.no_grad():
detections = self.model.predict(pil_image, threshold=confidence_threshold)
inference_time = time.time() - start_time
# Process results
results = []
if detections is not None and len(detections) > 0:
for i in range(len(detections)):
try:
class_id = int(detections.class_id[i])
if 0 <= class_id < len(self.class_names):
x1, y1, x2, y2 = detections.xyxy[i].tolist()
results.append({
'class_name': self.class_names[class_id],
'class_id': class_id,
'x1': float(x1),
'y1': float(y1),
'x2': float(x2),
'y2': float(y2),
'width': float(x2 - x1),
'height': float(y2 - y1),
'confidence': float(detections.confidence[i]),
'center_x': float((x1 + x2) / 2),
'center_y': float((y1 + y2) / 2),
'area': float((x2 - x1) * (y2 - y1))
})
except Exception as e:
print(f"⚠️ Error processing detection {i}: {e}")
continue
df = pd.DataFrame(results)
print(f"βœ… Found {len(df)} detections in {inference_time:.3f}s")
if len(df) > 0:
print("🎯 Detections:")
for class_name, count in df['class_name'].value_counts().items():
print(f" - {class_name}: {count}")
return df
def save_results(self, df: pd.DataFrame, output_path: str, format: str = 'csv', include_metadata: bool = True):
"""
Save DataFrame results to file with optional metadata.
Args:
df: DataFrame with detections
output_path: Output file path
format: 'csv', 'json', or 'parquet'
include_metadata: Whether to include model metadata
"""
output_path = Path(output_path)
output_path.parent.mkdir(parents=True, exist_ok=True)
if format.lower() == 'csv':
df.to_csv(output_path, index=False)
elif format.lower() == 'json':
result_data = {
'detections': df.to_dict('records'),
'summary': {
'total_detections': len(df),
'classes': df['class_name'].value_counts().to_dict() if len(df) > 0 else {}
}
}
if include_metadata:
result_data['model_info'] = self.model_info
with open(output_path, 'w') as f:
json.dump(result_data, f, indent=2)
elif format.lower() == 'parquet':
df.to_parquet(output_path, index=False)
else:
raise ValueError(f"Unsupported format: {format}. Use 'csv', 'json', or 'parquet'")
print(f"πŸ’Ύ Results saved to {output_path}")
def _save_video_results(self, df: pd.DataFrame, video_path: str, output_dir: Optional[str], stats: Dict):
"""Save video processing results with comprehensive metadata."""
if output_dir is None:
output_dir = os.path.dirname(video_path)
video_name = Path(video_path).stem
# Save main results
csv_path = Path(output_dir) / f"{video_name}_detections.csv"
json_path = Path(output_dir) / f"{video_name}_analysis.json"
# CSV with raw data
df.to_csv(csv_path, index=False)
# JSON with analysis
analysis = {
'video_info': {
'path': video_path,
'name': video_name,
**stats
},
'detection_summary': {
'total_detections': len(df),
'classes': df['class_name'].value_counts().to_dict() if len(df) > 0 else {},
'confidence_stats': {
'mean': float(df['confidence'].mean()) if len(df) > 0 else 0,
'median': float(df['confidence'].median()) if len(df) > 0 else 0,
'min': float(df['confidence'].min()) if len(df) > 0 else 0,
'max': float(df['confidence'].max()) if len(df) > 0 else 0
}
},
'model_info': self.model_info,
'detections': df.to_dict('records')
}
with open(json_path, 'w') as f:
json.dump(analysis, f, indent=2)
print(f"πŸ“Š Analysis saved:")
print(f" - CSV: {csv_path}")
print(f" - JSON: {json_path}")
def analyze_ball_possession(self, df: pd.DataFrame, distance_threshold: float = 100) -> pd.DataFrame:
"""
Analyze which players are near the ball (ball possession analysis).
Args:
df: DataFrame from process_video()
distance_threshold: Maximum distance to consider "near ball" (pixels)
Returns:
DataFrame with ball possession events
"""
print(f"⚽ Analyzing ball possession (threshold: {distance_threshold}px)")
ball_df = df[df['class_name'] == 'ball'].copy()
player_df = df[df['class_name'] == 'player'].copy()
possession_events = []
for frame in ball_df['frame'].unique():
ball_in_frame = ball_df[ball_df['frame'] == frame]
players_in_frame = player_df[player_df['frame'] == frame]
if len(ball_in_frame) > 0 and len(players_in_frame) > 0:
ball_center = ball_in_frame.iloc[0]
for _, player in players_in_frame.iterrows():
distance = np.sqrt(
(ball_center['center_x'] - player['center_x'])**2 +
(ball_center['center_y'] - player['center_y'])**2
)
if distance <= distance_threshold:
possession_events.append({
'frame': frame,
'timestamp': player['timestamp'],
'player_x': player['center_x'],
'player_y': player['center_y'],
'ball_x': ball_center['center_x'],
'ball_y': ball_center['center_y'],
'distance_to_ball': float(distance),
'ball_confidence': ball_center['confidence'],
'player_confidence': player['confidence']
})
possession_df = pd.DataFrame(possession_events)
if len(possession_df) > 0:
print(f"βœ… Found {len(possession_df)} possession events")
print(f"🎯 Average distance to ball: {possession_df['distance_to_ball'].mean():.1f}px")
else:
print("❌ No possession events found")
return possession_df
def get_model_info(self) -> Dict:
"""Get comprehensive model information."""
return {
**self.model_info,
'classes': self.class_names,
'device': self.device,
'checkpoint_path': str(self.model_path)
}
# Example usage and testing
if __name__ == "__main__":
# Initialize model
print("πŸš€ RF-DETR SoccerNet Demo")
model = RFDETRSoccerNet()
# Example with sample video (replace with your video path)
# df = model.process_video("sample_soccer_match.mp4", confidence_threshold=0.5)
# model.save_results(df, "match_analysis.json", format="json")
# Example ball possession analysis
# possession_df = model.analyze_ball_possession(df, distance_threshold=100)
# model.save_results(possession_df, "ball_possession.csv")
print("βœ… Demo complete! Replace with your video path to test.")