Lib/Core.py

import os
import sys
sys.path.append(os.getcwd())

import cv2
import numpy as np
import torch
from matplotlib import pyplot as plt
from ultralytics import YOLO
from Lib.Consts import LABELS, COLOR_MAP, COLOR_MAP_RGB


DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")


class CablePoleSegmentation():
    def __init__(self, model_path=None, retina_mask=False):
        if not model_path:
            model_path = "./weight/yolov9c-cable-seg.pt"
        self._RetinaMask=retina_mask
        self.Model = None
        self.PrepareModel(model_path)


    def PrepareModel(self, model_path):
        self.Model = YOLO(model_path)
        self.Model.fuse()


    def ScaleMasks(self, masks: torch.Tensor, shape: tuple) -> torch.Tensor:
        masks = masks.unsqueeze(0)
        interpolatedMask:torch.Tensor = torch.nn.functional.interpolate(masks, shape, mode="nearest")
        interpolatedMask = interpolatedMask.squeeze(0)
        return interpolatedMask
    

    def ParseResults(self, results, threshold=0.5, scale_masks=True):
        batches = []
        
        SCORES = torch.Tensor([]).to(DEVICE)
        CLASSES = torch.Tensor([]).to(DEVICE)
        MASKS = torch.Tensor([]).to(DEVICE)
        BOXES = torch.Tensor([]).to(DEVICE)

        with torch.no_grad():
            for result in results:
                original_shape = result.orig_shape
                _scores = result.boxes.conf        # 7
                _classes = result.boxes.cls         # 7
                _masks = result.masks.data         # 7, 480, 640
                _boxes = result.boxes.xyxy         # 7, 4
                
                # Threshold Filter
                conditions = _scores > threshold
                SCORES = torch.cat((SCORES, _scores[conditions]), dim=0)
                CLASSES = torch.cat((CLASSES, _classes[conditions]), dim=0)
                BOXES = torch.cat((BOXES, _boxes[conditions]), dim=0)
                mask = _masks[conditions]

                if mask.shape[0] == 0:
                    continue

                if scale_masks:
                    mask = self.ScaleMasks(mask, original_shape[:2])

                MASKS = torch.cat((MASKS, mask), dim=0)
                
                batches += [(SCORES, CLASSES, MASKS, BOXES)]

        return batches


    def DrawResults(self, image, scores: torch.Tensor, classes: torch.Tensor, masks: torch.Tensor, boxes: torch.Tensor, labels:dict=LABELS, class_filter:list=None):
        _image = np.array(image).copy()
        _image = cv2.cvtColor(_image, cv2.COLOR_BGR2RGB)
        maskCanvas = np.zeros_like(_image)


        with torch.no_grad():
            scores = scores.cpu().numpy()
            classes = classes.cpu().numpy().astype(np.int32)
            masks = masks.cpu().numpy()
            boxes = boxes.cpu().numpy()
            colors = list(COLOR_MAP_RGB.values())

        for score, cls, mask, box in zip(scores, classes, masks, boxes):
            label = labels[cls]
            _color = colors[cls]

            if class_filter and cls not in class_filter:
                continue

            box = box.astype(np.int32)
            mask = (cv2.cvtColor(mask, cv2.COLOR_GRAY2BGR)*_color).astype(np.uint8)
            maskCanvas = cv2.addWeighted(maskCanvas, 1.0, mask, 1.0, 0)
            maskCanvas = cv2.rectangle(maskCanvas, (box[0], box[1]), (box[2], box[3]), color=_color, thickness=5)  # Red color for bounding box
            maskCanvas = cv2.putText(maskCanvas, f"{label} : {score:.2f}", (box[0], box[1] - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.9, color=_color, thickness=2)
        
        canvas = cv2.addWeighted(_image, 1.0, maskCanvas.astype(np.uint8), 0.5, 0)
        return canvas, maskCanvas


    def Process(self, image, model_threshold=0.6, overall_threshold=0.6, iou=0.7,  class_filter:list=None):
        
        with torch.no_grad():
            results = self.Model(
                image,
                save=False,
                show_boxes=False,
                project="./inference/",
                conf=model_threshold,
                iou=iou,
                retina_masks=False,
                stream=True,
                classes=class_filter,
                device=DEVICE
            )

            batches = self.ParseResults(results, threshold=overall_threshold, scale_masks=True)

            return batches
            




if "__main__" == __name__:
    test = "data/DJI_20240905091530_0003_W.JPG"
    image = cv2.imread(test)
    model = CablePoleSegmentation(retina_mask=False)
    batches = model.Process(image)

    if len(batches) == 0:
        exit()

    scores, classes, masks, boxes = batches[0] # First
    canvas, mask = model.DrawResults(image, scores, classes, masks, boxes, class_filter=None)
    print(canvas.shape)


    #! Plot
    fig, axs = plt.subplots(1, 3, figsize=(27, 15))
    axs[0].imshow(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))
    axs[0].set_title("Orijinal Görüntü")

    axs[1].imshow(mask)
    axs[1].set_title("Segmentasyon Maskesi")

    axs[2].imshow(canvas)
    axs[2].set_title("Sonuç")

    plt.tight_layout()
    plt.show()