update

acf7a04 11 months ago

20.6 kB

from pathlib import Path
from typing import List, Tuple, Dict, Optional
import sys, os
sys.path.append(os.path.dirname(os.path.abspath(__file__)))
import onnxruntime as ort
import numpy as np
import cv2
from torchvision.ops import batched_nms
import torch
from ultralytics import YOLO
from numpy import ndarray
from pydantic import BaseModel
from team_cluster import TeamClassifier
from utils import (
    BoundingBox, 
    Constants,
    suppress_small_contained_boxes,
    classify_teams_batch,
)


class TVFrameResult(BaseModel):
    frame_id: int
    boxes: List[BoundingBox]
    keypoints: List[Tuple[int, int]]


class Miner:
    """
    Football video analysis system for object detection and team classification.
    """
    # Use constants from utils
    SMALL_CONTAINED_IOA = Constants.SMALL_CONTAINED_IOA
    SMALL_RATIO_MAX = Constants.SMALL_RATIO_MAX
    SINGLE_PLAYER_HUE_PIVOT = Constants.SINGLE_PLAYER_HUE_PIVOT
    CORNER_INDICES = Constants.CORNER_INDICES
    KEYPOINTS_CONFIDENCE = Constants.KEYPOINTS_CONFIDENCE
    CORNER_CONFIDENCE = Constants.CORNER_CONFIDENCE
    GOALKEEPER_POSITION_MARGIN = Constants.GOALKEEPER_POSITION_MARGIN
    MIN_SAMPLES_FOR_FIT = 16  # Minimum player crops needed before fitting TeamClassifier
    MAX_SAMPLES_FOR_FIT = 500  # Maximum samples to avoid overfitting

    def __init__(self, path_hf_repo: Path) -> None:
        providers = [
            'CUDAExecutionProvider',
            'CPUExecutionProvider'
        ]
        model_path = path_hf_repo / "detection.onnx"
        session = ort.InferenceSession(model_path, providers=providers)

        input_name = session.get_inputs()[0].name
        height = width = 640
        dummy = np.zeros((1, 3, height, width), dtype=np.float32)
        session.run(None, {input_name: dummy})
        model = session
        self.bbox_model = model
        
        print("BBox Model Loaded")
        self.keypoints_model = YOLO(path_hf_repo / "keypoint.pt")
        print("Keypoints Model (keypoint.pt) Loaded")
        # Initialize team classifier with OSNet model
        team_model_path = path_hf_repo / "osnet_model.pth.tar-100"
        device = 'cuda'
        self.team_classifier = TeamClassifier(
            device=device,
            batch_size=32,
            model_name=str(team_model_path)
        )
        print("Team Classifier Loaded")
        
        # Team classification state
        self.team_classifier_fitted = False
        self.player_crops_for_fit = []  # Collect samples across frames

    def __repr__(self) -> str:
        return (
            f"BBox Model: {type(self.bbox_model).__name__}\n"
            f"Keypoints Model: {type(self.keypoints_model).__name__}"
        )



    def _handle_multiple_goalkeepers(self, boxes: List[BoundingBox]) -> List[BoundingBox]:
        """
        Handle goalkeeper detection issues:
        1. Fix misplaced goalkeepers (standing in middle of field)
        2. Limit to maximum 2 goalkeepers (one from each team)
        
        Returns:
            Filtered list of boxes with corrected goalkeepers
        """
        # Step 1: Fix misplaced goalkeepers first
        # Convert goalkeepers in middle of field to regular players
        boxes = self._fix_misplaced_goalkeepers(boxes)
        
        # Step 2: Handle multiple goalkeepers (after fixing misplaced ones)
        gk_idxs = [i for i, bb in enumerate(boxes) if int(bb.cls_id) == 1]
        if len(gk_idxs) <= 2:
            return boxes
            
        # Sort goalkeepers by confidence (highest first)
        gk_idxs_sorted = sorted(gk_idxs, key=lambda i: boxes[i].conf, reverse=True)
        keep_gk_idxs = set(gk_idxs_sorted[:2])  # Keep top 2 goalkeepers
        
        # Create new list keeping only top 2 goalkeepers
        filtered_boxes = []
        for i, box in enumerate(boxes):
            if int(box.cls_id) == 1:
                # Only keep the top 2 goalkeepers by confidence
                if i in keep_gk_idxs:
                    filtered_boxes.append(box)
                # Skip extra goalkeepers
            else:
                # Keep all non-goalkeeper boxes
                filtered_boxes.append(box)
        
        return filtered_boxes

    def _fix_misplaced_goalkeepers(self, boxes: List[BoundingBox]) -> List[BoundingBox]:
        """
        """
        gk_idxs = [i for i, bb in enumerate(boxes) if int(bb.cls_id) == 1]
        player_idxs = [i for i, bb in enumerate(boxes) if int(bb.cls_id) == 2]
        
        if len(gk_idxs) == 0 or len(player_idxs) < 2:
            return boxes
            
        updated_boxes = boxes.copy()
        
        for gk_idx in gk_idxs:
            if boxes[gk_idx].conf < 0.3:
                updated_boxes[gk_idx].cls_id = 2
                
        return updated_boxes


    def _pre_process_img(self, frames: List[np.ndarray], scale: float = 640.0) -> np.ndarray:
        """
        Preprocess images for ONNX inference.
        
        Args:
            frames: List of BGR frames
            scale: Target scale for resizing
            
        Returns:
            Preprocessed numpy array ready for ONNX inference
        """
        imgs = np.stack([cv2.resize(frame, (int(scale), int(scale))) for frame in frames])
        imgs = imgs.transpose(0, 3, 1, 2)  # BHWC to BCHW
        imgs = imgs.astype(np.float32) / 255.0  # Normalize to [0, 1]
        return imgs

    def _post_process_output(self, outputs: np.ndarray, x_scale: float, y_scale: float, 
                            conf_thresh: float = 0.6, nms_thresh: float = 0.55) -> List[List[Tuple]]:
        """
        Post-process ONNX model outputs to get detections.
        
        Args:
            outputs: Raw ONNX model outputs
            x_scale: X-axis scaling factor 
            y_scale: Y-axis scaling factor
            conf_thresh: Confidence threshold
            nms_thresh: NMS threshold
            
        Returns:
            List of detections for each frame: [(box, conf, class_id), ...]
        """
        B, C, N = outputs.shape
        outputs = torch.from_numpy(outputs)
        outputs = outputs.permute(0, 2, 1)  # B,C,N -> B,N,C
        
        boxes = outputs[..., :4]
        class_scores = 1 / (1 + torch.exp(-outputs[..., 4:]))  # Sigmoid activation
        conf, class_id = class_scores.max(dim=2)

        mask = conf > conf_thresh
        
        # Special handling for balls - keep best one even with lower confidence
        for i in range(class_id.shape[0]):  # loop over batch
            # Find detections that are balls
            ball_mask = class_id[i] == 0
            ball_idx = ball_mask.nonzero(as_tuple=True)[0]
            if ball_idx.numel() > 0:
                # Pick the one with the highest confidence
                best_ball_idx = ball_idx[conf[i, ball_idx].argmax()]
                if conf[i, best_ball_idx] >= 0.55:  # apply confidence threshold
                    mask[i, best_ball_idx] = True
        
        batch_idx, pred_idx = mask.nonzero(as_tuple=True)

        if len(batch_idx) == 0:
            return [[] for _ in range(B)]
        
        boxes = boxes[batch_idx, pred_idx]
        conf = conf[batch_idx, pred_idx]
        class_id = class_id[batch_idx, pred_idx]

        # Convert from center format to xyxy format
        x, y, w, h = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
        x1 = (x - w / 2) * x_scale
        y1 = (y - h / 2) * y_scale
        x2 = (x + w / 2) * x_scale
        y2 = (y + h / 2) * y_scale
        boxes_xyxy = torch.stack([x1, y1, x2, y2], dim=1)

        # Apply batched NMS
        max_coord = 1e4
        offset = batch_idx.to(boxes_xyxy) * max_coord
        boxes_for_nms = boxes_xyxy + offset[:, None]

        keep = batched_nms(boxes_for_nms, conf, batch_idx, nms_thresh)

        boxes_final = boxes_xyxy[keep]
        conf_final = conf[keep]
        class_final = class_id[keep]
        batch_final = batch_idx[keep]

        # Group results by batch
        results = [[] for _ in range(B)]
        for b in range(B):
            mask_b = batch_final == b
            if mask_b.sum() == 0:
                continue
            results[b] = list(zip(boxes_final[mask_b].numpy(),
                                  conf_final[mask_b].numpy(),
                                  class_final[mask_b].numpy()))
        return results

    def _ioa(self, a: BoundingBox, b: BoundingBox) -> float:
        inter = self._intersect_area(a, b)
        aa = self._area(a)
        if aa <= 0:
            return 0.0
        return inter / aa

    def suppress_small_contained(self, boxes: List[BoundingBox]) -> List[BoundingBox]:
        if len(boxes) <= 1:
            return boxes
        keep = [True] * len(boxes)
        areas = [self._area(bb) for bb in boxes]
        for i in range(len(boxes)):
            if not keep[i]:
                continue
            for j in range(len(boxes)):
                if i == j or not keep[j]:
                    continue
                ai, aj = areas[i], areas[j]
                if ai == 0 or aj == 0:
                    continue
                if ai <= aj:
                    ratio = ai / aj
                    if ratio <= self.SMALL_RATIO_MAX:
                        ioa_i_in_j = self._ioa(boxes[i], boxes[j])
                        if ioa_i_in_j >= self.SMALL_CONTAINED_IOA:
                            keep[i] = False
                            break
                else:
                    ratio = aj / ai
                    if ratio <= self.SMALL_RATIO_MAX:
                        ioa_j_in_i = self._ioa(boxes[j], boxes[i])
                        if ioa_j_in_i >= self.SMALL_CONTAINED_IOA:
                            keep[j] = False
        return [bb for bb, k in zip(boxes, keep) if k]

    def _detect_objects_batch(self, batch_images: List[ndarray], offset: int) -> Dict[int, List[BoundingBox]]:
        """
        Phase 1: Object detection for all frames in batch.
        Returns detected objects with players still having class_id=2 (before team classification).
        
        Args:
            batch_images: List of images to process
            offset: Frame offset for numbering
            
        Returns:
            Dictionary mapping frame_id to list of detected boxes
        """
        bboxes: Dict[int, List[BoundingBox]] = {}

        if len(batch_images) == 0:
            return bboxes
            
        print(f"Processing batch of {len(batch_images)} images")
        
        # Get original image dimensions for scaling
        height, width = batch_images[0].shape[:2]
        scale = 640.0
        x_scale = width / scale
        y_scale = height / scale
        
        # Memory optimization: Process smaller batches if needed
        max_batch_size = 32  # Reduce batch size further to prevent memory issues
        if len(batch_images) > max_batch_size:
            print(f"Large batch detected ({len(batch_images)} images), splitting into smaller batches of {max_batch_size}")
            # Process in smaller chunks
            all_bboxes = {}
            for chunk_start in range(0, len(batch_images), max_batch_size):
                chunk_end = min(chunk_start + max_batch_size, len(batch_images))
                chunk_images = batch_images[chunk_start:chunk_end]
                chunk_offset = offset + chunk_start
                print(f"Processing chunk {chunk_start//max_batch_size + 1}: images {chunk_start}-{chunk_end-1}")
                chunk_bboxes = self._detect_objects_batch(chunk_images, chunk_offset)
                all_bboxes.update(chunk_bboxes)
            return all_bboxes
        
        # Preprocess images for ONNX inference
        imgs = self._pre_process_img(batch_images, scale)
        actual_batch_size = len(batch_images)
        
        # Handle batch size mismatch - pad if needed
        model_batch_size = self.bbox_model.get_inputs()[0].shape[0]
        print(f"Model input shape: {self.bbox_model.get_inputs()[0].shape}, batch_size: {model_batch_size}")
        
        if model_batch_size is not None:
            try:
                # Handle dynamic batch size (None, -1, 'None')
                if str(model_batch_size) in ['None', '-1'] or model_batch_size == -1:
                    model_batch_size = None
                else:
                    model_batch_size = int(model_batch_size)
            except (ValueError, TypeError):
                model_batch_size = None
        
        print(f"Processed model_batch_size: {model_batch_size}, actual_batch_size: {actual_batch_size}")
                
        if model_batch_size and actual_batch_size < model_batch_size:
            padding_size = model_batch_size - actual_batch_size
            dummy_img = np.zeros((1, 3, int(scale), int(scale)), dtype=np.float32)
            padding = np.repeat(dummy_img, padding_size, axis=0)
            imgs = np.vstack([imgs, padding])
        
        # ONNX inference with error handling
        try:
            input_name = self.bbox_model.get_inputs()[0].name
            import time
            start_time = time.time()
            outputs = self.bbox_model.run(None, {input_name: imgs})[0]
            inference_time = time.time() - start_time
            print(f"Inference time: {inference_time:.3f}s for {actual_batch_size} images")
            
            # Remove padded results if we added padding
            if model_batch_size and isinstance(model_batch_size, int) and actual_batch_size < model_batch_size:
                outputs = outputs[:actual_batch_size]
            
            # Post-process outputs to get detections
            raw_results = self._post_process_output(np.array(outputs), x_scale, y_scale)
            
        except Exception as e:
            print(f"Error during ONNX inference: {e}")
            return bboxes
        
        if not raw_results:
            return bboxes
            
        # Convert raw results to BoundingBox objects and apply processing
        for frame_idx_in_batch, frame_detections in enumerate(raw_results):
            if not frame_detections:
                continue
                
            # Convert to BoundingBox objects
            boxes: List[BoundingBox] = []
            for box, conf, cls_id in frame_detections:
                x1, y1, x2, y2 = box
                if int(cls_id) < 4:
                    boxes.append(
                        BoundingBox(
                            x1=int(x1),
                            y1=int(y1),
                            x2=int(x2),
                            y2=int(y2),
                            cls_id=int(cls_id),
                            conf=float(conf),
                        )
                    )
            
            # Handle footballs - keep only the best one
            footballs = [bb for bb in boxes if int(bb.cls_id) == 0]
            if len(footballs) > 1:
                best_ball = max(footballs, key=lambda b: b.conf)
                boxes = [bb for bb in boxes if int(bb.cls_id) != 0]
                boxes.append(best_ball)

            # Remove overlapping small boxes
            boxes = suppress_small_contained_boxes(boxes, self.SMALL_CONTAINED_IOA, self.SMALL_RATIO_MAX)
            
            # Handle goalkeeper detection issues:
            # 1. Fix misplaced goalkeepers (convert to players if standing in middle)
            # 2. Allow up to 2 goalkeepers maximum (one from each team)
            # Goalkeepers remain class_id = 1 (no team assignment)
            boxes = self._handle_multiple_goalkeepers(boxes)
            
            # Store results (players still have class_id=2, will be classified in phase 2)
            frame_id = offset + frame_idx_in_batch
            bboxes[frame_id] = boxes
            
        return bboxes


    def predict_batch(
        self,
        batch_images: List[ndarray],
        offset: int,
        n_keypoints: int,
        task_type: Optional[str] = None,
    ) -> List[TVFrameResult]:
        process_objects = task_type is None or task_type == "object"
        process_keypoints = task_type is None or task_type == "keypoint"
        
        # Phase 1: Object Detection for all frames
        bboxes: Dict[int, List[BoundingBox]] = {}
        if process_objects:
            bboxes = self._detect_objects_batch(batch_images, offset)
        
        import time
        time_start = time.time()
        # Phase 2: Team Classification for all detected players
        if process_objects and bboxes:
            bboxes, self.team_classifier_fitted, self.player_crops_for_fit = classify_teams_batch(
                self.team_classifier, 
                self.team_classifier_fitted,
                self.player_crops_for_fit,
                batch_images, 
                bboxes, 
                offset,
                self.MIN_SAMPLES_FOR_FIT,
                self.MAX_SAMPLES_FOR_FIT,
                self.SINGLE_PLAYER_HUE_PIVOT
            )
        self.team_classifier_fitted = False
        self.player_crops_for_fit = []
        print(f"Time Team Classification: {time.time() - time_start} s")

        # Phase 3: Keypoint Detection
        keypoints: Dict[int, List[Tuple[int, int]]] = {}
        if process_keypoints:
            keypoints = self._detect_keypoints_batch(batch_images, offset, n_keypoints)
        
        # Phase 4: Combine results
        results: List[TVFrameResult] = []
        for frame_number in range(offset, offset + len(batch_images)):
            results.append(
                TVFrameResult(
                    frame_id=frame_number,
                    boxes=bboxes.get(frame_number, []),
                    keypoints=keypoints.get(
                        frame_number,
                        [(0, 0) for _ in range(n_keypoints)],
                    ),
                )
            )
        return results

    def _detect_keypoints_batch(self, batch_images: List[ndarray], 
                               offset: int, n_keypoints: int) -> Dict[int, List[Tuple[int, int]]]:
        """
        Phase 3: Keypoint detection for all frames in batch.
        
        Args:
            batch_images: List of images to process
            offset: Frame offset for numbering  
            n_keypoints: Number of keypoints expected
            
        Returns:
            Dictionary mapping frame_id to list of keypoint coordinates
        """
        keypoints: Dict[int, List[Tuple[int, int]]] = {}
        keypoints_model_results = self.keypoints_model.predict(batch_images)
        
        if keypoints_model_results is None:
            return keypoints
            
        for frame_idx_in_batch, detection in enumerate(keypoints_model_results):
            if not hasattr(detection, "keypoints") or detection.keypoints is None:
                continue
                
            # Extract keypoints with confidence
            frame_keypoints_with_conf: List[Tuple[int, int, float]] = []
            for i, part_points in enumerate(detection.keypoints.data):
                for k_id, (x, y, _) in enumerate(part_points):
                    confidence = float(detection.keypoints.conf[i][k_id])
                    frame_keypoints_with_conf.append((int(x), int(y), confidence))
            
            # Pad or truncate to expected number of keypoints
            if len(frame_keypoints_with_conf) < n_keypoints:
                frame_keypoints_with_conf.extend(
                    [(0, 0, 0.0)] * (n_keypoints - len(frame_keypoints_with_conf))
                )
            else:
                frame_keypoints_with_conf = frame_keypoints_with_conf[:n_keypoints]
            
            # Filter keypoints based on confidence thresholds
            filtered_keypoints: List[Tuple[int, int]] = []
            for idx, (x, y, confidence) in enumerate(frame_keypoints_with_conf):
                if idx in self.CORNER_INDICES:
                    # Corner keypoints have lower confidence threshold
                    if confidence < 0.3:
                        filtered_keypoints.append((0, 0))
                    else:
                        filtered_keypoints.append((int(x), int(y)))
                else:
                    # Regular keypoints  
                    if confidence < 0.5:
                        filtered_keypoints.append((0, 0))
                    else:
                        filtered_keypoints.append((int(x), int(y)))
            
            frame_id = offset + frame_idx_in_batch
            keypoints[frame_id] = filtered_keypoints
            
        return keypoints

	from pathlib import Path
	from typing import List, Tuple, Dict, Optional
	import sys, os
	sys.path.append(os.path.dirname(os.path.abspath(__file__)))
	import onnxruntime as ort
	import numpy as np
	import cv2
	from torchvision.ops import batched_nms
	import torch
	from ultralytics import YOLO
	from numpy import ndarray
	from pydantic import BaseModel
	from team_cluster import TeamClassifier
	from utils import (
	BoundingBox,
	Constants,
	suppress_small_contained_boxes,
	classify_teams_batch,
	)


	class TVFrameResult(BaseModel):
	frame_id: int
	boxes: List[BoundingBox]
	keypoints: List[Tuple[int, int]]


	class Miner:
	"""
	Football video analysis system for object detection and team classification.
	"""
	# Use constants from utils
	SMALL_CONTAINED_IOA = Constants.SMALL_CONTAINED_IOA
	SMALL_RATIO_MAX = Constants.SMALL_RATIO_MAX
	SINGLE_PLAYER_HUE_PIVOT = Constants.SINGLE_PLAYER_HUE_PIVOT
	CORNER_INDICES = Constants.CORNER_INDICES
	KEYPOINTS_CONFIDENCE = Constants.KEYPOINTS_CONFIDENCE
	CORNER_CONFIDENCE = Constants.CORNER_CONFIDENCE
	GOALKEEPER_POSITION_MARGIN = Constants.GOALKEEPER_POSITION_MARGIN
	MIN_SAMPLES_FOR_FIT = 16 # Minimum player crops needed before fitting TeamClassifier
	MAX_SAMPLES_FOR_FIT = 500 # Maximum samples to avoid overfitting

	def __init__(self, path_hf_repo: Path) -> None:
	providers = [
	'CUDAExecutionProvider',
	'CPUExecutionProvider'
	]
	model_path = path_hf_repo / "detection.onnx"
	session = ort.InferenceSession(model_path, providers=providers)

	input_name = session.get_inputs()[0].name
	height = width = 640
	dummy = np.zeros((1, 3, height, width), dtype=np.float32)
	session.run(None, {input_name: dummy})
	model = session
	self.bbox_model = model

	print("BBox Model Loaded")
	self.keypoints_model = YOLO(path_hf_repo / "keypoint.pt")
	print("Keypoints Model (keypoint.pt) Loaded")
	# Initialize team classifier with OSNet model
	team_model_path = path_hf_repo / "osnet_model.pth.tar-100"
	device = 'cuda'
	self.team_classifier = TeamClassifier(
	device=device,
	batch_size=32,
	model_name=str(team_model_path)
	)
	print("Team Classifier Loaded")

	# Team classification state
	self.team_classifier_fitted = False
	self.player_crops_for_fit = [] # Collect samples across frames

	def __repr__(self) -> str:
	return (
	f"BBox Model: {type(self.bbox_model).__name__}\n"
	f"Keypoints Model: {type(self.keypoints_model).__name__}"
	)



	def _handle_multiple_goalkeepers(self, boxes: List[BoundingBox]) -> List[BoundingBox]:
	"""
	Handle goalkeeper detection issues:
	1. Fix misplaced goalkeepers (standing in middle of field)
	2. Limit to maximum 2 goalkeepers (one from each team)

	Returns:
	Filtered list of boxes with corrected goalkeepers
	"""
	# Step 1: Fix misplaced goalkeepers first
	# Convert goalkeepers in middle of field to regular players
	boxes = self._fix_misplaced_goalkeepers(boxes)

	# Step 2: Handle multiple goalkeepers (after fixing misplaced ones)
	gk_idxs = [i for i, bb in enumerate(boxes) if int(bb.cls_id) == 1]
	if len(gk_idxs) <= 2:
	return boxes

	# Sort goalkeepers by confidence (highest first)
	gk_idxs_sorted = sorted(gk_idxs, key=lambda i: boxes[i].conf, reverse=True)
	keep_gk_idxs = set(gk_idxs_sorted[:2]) # Keep top 2 goalkeepers

	# Create new list keeping only top 2 goalkeepers
	filtered_boxes = []
	for i, box in enumerate(boxes):
	if int(box.cls_id) == 1:
	# Only keep the top 2 goalkeepers by confidence
	if i in keep_gk_idxs:
	filtered_boxes.append(box)
	# Skip extra goalkeepers
	else:
	# Keep all non-goalkeeper boxes
	filtered_boxes.append(box)

	return filtered_boxes

	def _fix_misplaced_goalkeepers(self, boxes: List[BoundingBox]) -> List[BoundingBox]:
	"""
	"""
	gk_idxs = [i for i, bb in enumerate(boxes) if int(bb.cls_id) == 1]
	player_idxs = [i for i, bb in enumerate(boxes) if int(bb.cls_id) == 2]

	if len(gk_idxs) == 0 or len(player_idxs) < 2:
	return boxes

	updated_boxes = boxes.copy()

	for gk_idx in gk_idxs:
	if boxes[gk_idx].conf < 0.3:
	updated_boxes[gk_idx].cls_id = 2

	return updated_boxes


	def _pre_process_img(self, frames: List[np.ndarray], scale: float = 640.0) -> np.ndarray:
	"""
	Preprocess images for ONNX inference.

	Args:
	frames: List of BGR frames
	scale: Target scale for resizing

	Returns:
	Preprocessed numpy array ready for ONNX inference
	"""
	imgs = np.stack([cv2.resize(frame, (int(scale), int(scale))) for frame in frames])
	imgs = imgs.transpose(0, 3, 1, 2) # BHWC to BCHW
	imgs = imgs.astype(np.float32) / 255.0 # Normalize to [0, 1]
	return imgs

	def _post_process_output(self, outputs: np.ndarray, x_scale: float, y_scale: float,
	conf_thresh: float = 0.6, nms_thresh: float = 0.55) -> List[List[Tuple]]:
	"""
	Post-process ONNX model outputs to get detections.

	Args:
	outputs: Raw ONNX model outputs
	x_scale: X-axis scaling factor
	y_scale: Y-axis scaling factor
	conf_thresh: Confidence threshold
	nms_thresh: NMS threshold

	Returns:
	List of detections for each frame: [(box, conf, class_id), ...]
	"""
	B, C, N = outputs.shape
	outputs = torch.from_numpy(outputs)
	outputs = outputs.permute(0, 2, 1) # B,C,N -> B,N,C

	boxes = outputs[..., :4]
	class_scores = 1 / (1 + torch.exp(-outputs[..., 4:])) # Sigmoid activation
	conf, class_id = class_scores.max(dim=2)

	mask = conf > conf_thresh

	# Special handling for balls - keep best one even with lower confidence
	for i in range(class_id.shape[0]): # loop over batch
	# Find detections that are balls
	ball_mask = class_id[i] == 0
	ball_idx = ball_mask.nonzero(as_tuple=True)[0]
	if ball_idx.numel() > 0:
	# Pick the one with the highest confidence
	best_ball_idx = ball_idx[conf[i, ball_idx].argmax()]
	if conf[i, best_ball_idx] >= 0.55: # apply confidence threshold
	mask[i, best_ball_idx] = True

	batch_idx, pred_idx = mask.nonzero(as_tuple=True)

	if len(batch_idx) == 0:
	return [[] for _ in range(B)]

	boxes = boxes[batch_idx, pred_idx]
	conf = conf[batch_idx, pred_idx]
	class_id = class_id[batch_idx, pred_idx]

	# Convert from center format to xyxy format
	x, y, w, h = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
	x1 = (x - w / 2) * x_scale
	y1 = (y - h / 2) * y_scale
	x2 = (x + w / 2) * x_scale
	y2 = (y + h / 2) * y_scale
	boxes_xyxy = torch.stack([x1, y1, x2, y2], dim=1)

	# Apply batched NMS
	max_coord = 1e4
	offset = batch_idx.to(boxes_xyxy) * max_coord
	boxes_for_nms = boxes_xyxy + offset[:, None]

	keep = batched_nms(boxes_for_nms, conf, batch_idx, nms_thresh)

	boxes_final = boxes_xyxy[keep]
	conf_final = conf[keep]
	class_final = class_id[keep]
	batch_final = batch_idx[keep]

	# Group results by batch
	results = [[] for _ in range(B)]
	for b in range(B):
	mask_b = batch_final == b
	if mask_b.sum() == 0:
	continue
	results[b] = list(zip(boxes_final[mask_b].numpy(),
	conf_final[mask_b].numpy(),
	class_final[mask_b].numpy()))
	return results

	def _ioa(self, a: BoundingBox, b: BoundingBox) -> float:
	inter = self._intersect_area(a, b)
	aa = self._area(a)
	if aa <= 0:
	return 0.0
	return inter / aa

	def suppress_small_contained(self, boxes: List[BoundingBox]) -> List[BoundingBox]:
	if len(boxes) <= 1:
	return boxes
	keep = [True] * len(boxes)
	areas = [self._area(bb) for bb in boxes]
	for i in range(len(boxes)):
	if not keep[i]:
	continue
	for j in range(len(boxes)):
	if i == j or not keep[j]:
	continue
	ai, aj = areas[i], areas[j]
	if ai == 0 or aj == 0:
	continue
	if ai <= aj:
	ratio = ai / aj
	if ratio <= self.SMALL_RATIO_MAX:
	ioa_i_in_j = self._ioa(boxes[i], boxes[j])
	if ioa_i_in_j >= self.SMALL_CONTAINED_IOA:
	keep[i] = False
	break
	else:
	ratio = aj / ai
	if ratio <= self.SMALL_RATIO_MAX:
	ioa_j_in_i = self._ioa(boxes[j], boxes[i])
	if ioa_j_in_i >= self.SMALL_CONTAINED_IOA:
	keep[j] = False
	return [bb for bb, k in zip(boxes, keep) if k]

	def _detect_objects_batch(self, batch_images: List[ndarray], offset: int) -> Dict[int, List[BoundingBox]]:
	"""
	Phase 1: Object detection for all frames in batch.
	Returns detected objects with players still having class_id=2 (before team classification).

	Args:
	batch_images: List of images to process
	offset: Frame offset for numbering

	Returns:
	Dictionary mapping frame_id to list of detected boxes
	"""
	bboxes: Dict[int, List[BoundingBox]] = {}

	if len(batch_images) == 0:
	return bboxes

	print(f"Processing batch of {len(batch_images)} images")

	# Get original image dimensions for scaling
	height, width = batch_images[0].shape[:2]
	scale = 640.0
	x_scale = width / scale
	y_scale = height / scale

	# Memory optimization: Process smaller batches if needed
	max_batch_size = 32 # Reduce batch size further to prevent memory issues
	if len(batch_images) > max_batch_size:
	print(f"Large batch detected ({len(batch_images)} images), splitting into smaller batches of {max_batch_size}")
	# Process in smaller chunks
	all_bboxes = {}
	for chunk_start in range(0, len(batch_images), max_batch_size):
	chunk_end = min(chunk_start + max_batch_size, len(batch_images))
	chunk_images = batch_images[chunk_start:chunk_end]
	chunk_offset = offset + chunk_start
	print(f"Processing chunk {chunk_start//max_batch_size + 1}: images {chunk_start}-{chunk_end-1}")
	chunk_bboxes = self._detect_objects_batch(chunk_images, chunk_offset)
	all_bboxes.update(chunk_bboxes)
	return all_bboxes

	# Preprocess images for ONNX inference
	imgs = self._pre_process_img(batch_images, scale)
	actual_batch_size = len(batch_images)

	# Handle batch size mismatch - pad if needed
	model_batch_size = self.bbox_model.get_inputs()[0].shape[0]
	print(f"Model input shape: {self.bbox_model.get_inputs()[0].shape}, batch_size: {model_batch_size}")

	if model_batch_size is not None:
	try:
	# Handle dynamic batch size (None, -1, 'None')
	if str(model_batch_size) in ['None', '-1'] or model_batch_size == -1:
	model_batch_size = None
	else:
	model_batch_size = int(model_batch_size)
	except (ValueError, TypeError):
	model_batch_size = None

	print(f"Processed model_batch_size: {model_batch_size}, actual_batch_size: {actual_batch_size}")

	if model_batch_size and actual_batch_size < model_batch_size:
	padding_size = model_batch_size - actual_batch_size
	dummy_img = np.zeros((1, 3, int(scale), int(scale)), dtype=np.float32)
	padding = np.repeat(dummy_img, padding_size, axis=0)
	imgs = np.vstack([imgs, padding])

	# ONNX inference with error handling
	try:
	input_name = self.bbox_model.get_inputs()[0].name
	import time
	start_time = time.time()
	outputs = self.bbox_model.run(None, {input_name: imgs})[0]
	inference_time = time.time() - start_time
	print(f"Inference time: {inference_time:.3f}s for {actual_batch_size} images")

	# Remove padded results if we added padding
	if model_batch_size and isinstance(model_batch_size, int) and actual_batch_size < model_batch_size:
	outputs = outputs[:actual_batch_size]

	# Post-process outputs to get detections
	raw_results = self._post_process_output(np.array(outputs), x_scale, y_scale)

	except Exception as e:
	print(f"Error during ONNX inference: {e}")
	return bboxes

	if not raw_results:
	return bboxes

	# Convert raw results to BoundingBox objects and apply processing
	for frame_idx_in_batch, frame_detections in enumerate(raw_results):
	if not frame_detections:
	continue

	# Convert to BoundingBox objects
	boxes: List[BoundingBox] = []
	for box, conf, cls_id in frame_detections:
	x1, y1, x2, y2 = box
	if int(cls_id) < 4:
	boxes.append(
	BoundingBox(
	x1=int(x1),
	y1=int(y1),
	x2=int(x2),
	y2=int(y2),
	cls_id=int(cls_id),
	conf=float(conf),
	)
	)

	# Handle footballs - keep only the best one
	footballs = [bb for bb in boxes if int(bb.cls_id) == 0]
	if len(footballs) > 1:
	best_ball = max(footballs, key=lambda b: b.conf)
	boxes = [bb for bb in boxes if int(bb.cls_id) != 0]
	boxes.append(best_ball)

	# Remove overlapping small boxes
	boxes = suppress_small_contained_boxes(boxes, self.SMALL_CONTAINED_IOA, self.SMALL_RATIO_MAX)

	# Handle goalkeeper detection issues:
	# 1. Fix misplaced goalkeepers (convert to players if standing in middle)
	# 2. Allow up to 2 goalkeepers maximum (one from each team)
	# Goalkeepers remain class_id = 1 (no team assignment)
	boxes = self._handle_multiple_goalkeepers(boxes)

	# Store results (players still have class_id=2, will be classified in phase 2)
	frame_id = offset + frame_idx_in_batch
	bboxes[frame_id] = boxes

	return bboxes


	def predict_batch(
	self,
	batch_images: List[ndarray],
	offset: int,
	n_keypoints: int,
	task_type: Optional[str] = None,
	) -> List[TVFrameResult]:
	process_objects = task_type is None or task_type == "object"
	process_keypoints = task_type is None or task_type == "keypoint"

	# Phase 1: Object Detection for all frames
	bboxes: Dict[int, List[BoundingBox]] = {}
	if process_objects:
	bboxes = self._detect_objects_batch(batch_images, offset)

	import time
	time_start = time.time()
	# Phase 2: Team Classification for all detected players
	if process_objects and bboxes:
	bboxes, self.team_classifier_fitted, self.player_crops_for_fit = classify_teams_batch(
	self.team_classifier,
	self.team_classifier_fitted,
	self.player_crops_for_fit,
	batch_images,
	bboxes,
	offset,
	self.MIN_SAMPLES_FOR_FIT,
	self.MAX_SAMPLES_FOR_FIT,
	self.SINGLE_PLAYER_HUE_PIVOT
	)
	self.team_classifier_fitted = False
	self.player_crops_for_fit = []
	print(f"Time Team Classification: {time.time() - time_start} s")

	# Phase 3: Keypoint Detection
	keypoints: Dict[int, List[Tuple[int, int]]] = {}
	if process_keypoints:
	keypoints = self._detect_keypoints_batch(batch_images, offset, n_keypoints)

	# Phase 4: Combine results
	results: List[TVFrameResult] = []
	for frame_number in range(offset, offset + len(batch_images)):
	results.append(
	TVFrameResult(
	frame_id=frame_number,
	boxes=bboxes.get(frame_number, []),
	keypoints=keypoints.get(
	frame_number,
	[(0, 0) for _ in range(n_keypoints)],
	),
	)
	)
	return results

	def _detect_keypoints_batch(self, batch_images: List[ndarray],
	offset: int, n_keypoints: int) -> Dict[int, List[Tuple[int, int]]]:
	"""
	Phase 3: Keypoint detection for all frames in batch.

	Args:
	batch_images: List of images to process
	offset: Frame offset for numbering
	n_keypoints: Number of keypoints expected

	Returns:
	Dictionary mapping frame_id to list of keypoint coordinates
	"""
	keypoints: Dict[int, List[Tuple[int, int]]] = {}
	keypoints_model_results = self.keypoints_model.predict(batch_images)

	if keypoints_model_results is None:
	return keypoints

	for frame_idx_in_batch, detection in enumerate(keypoints_model_results):
	if not hasattr(detection, "keypoints") or detection.keypoints is None:
	continue

	# Extract keypoints with confidence
	frame_keypoints_with_conf: List[Tuple[int, int, float]] = []
	for i, part_points in enumerate(detection.keypoints.data):
	for k_id, (x, y, _) in enumerate(part_points):
	confidence = float(detection.keypoints.conf[i][k_id])
	frame_keypoints_with_conf.append((int(x), int(y), confidence))

	# Pad or truncate to expected number of keypoints
	if len(frame_keypoints_with_conf) < n_keypoints:
	frame_keypoints_with_conf.extend(
	[(0, 0, 0.0)] * (n_keypoints - len(frame_keypoints_with_conf))
	)
	else:
	frame_keypoints_with_conf = frame_keypoints_with_conf[:n_keypoints]

	# Filter keypoints based on confidence thresholds
	filtered_keypoints: List[Tuple[int, int]] = []
	for idx, (x, y, confidence) in enumerate(frame_keypoints_with_conf):
	if idx in self.CORNER_INDICES:
	# Corner keypoints have lower confidence threshold
	if confidence < 0.3:
	filtered_keypoints.append((0, 0))
	else:
	filtered_keypoints.append((int(x), int(y)))
	else:
	# Regular keypoints
	if confidence < 0.5:
	filtered_keypoints.append((0, 0))
	else:
	filtered_keypoints.append((int(x), int(y)))

	frame_id = offset + frame_idx_in_batch
	keypoints[frame_id] = filtered_keypoints

	return keypoints