How to exchange Mask R-CNN from YOLOv3

[szk-ryoma]

I was able to exchange to the MASK model and run deepsort.py, but the bounding box appears very large and the accuracy is very low.
I know that if it is a still image it detects with very high accuracy.
So the model is well trained.

Please give me some ideas on how to solve this problem.

The following code is the detector.py for MASK R-CNN.

class MASK_RCNN(object):
def init(self, model_path, mask_thresh, bbox_thresh):
self.model = torch.load(model_path)
self.device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
self.model.eval()
self.mask_thresh = mask_thresh
self.bbox_thresh = bbox_thresh

def __call__(self, input_img):
    img = Image.open(ori_img)
    img = img.convert("RGB")
    
    transform = transforms.Compose([transforms.ToTensor()])
    img = transform(img)

    with torch.no_grad():  
        result = self.model([img.to(self.device)])

    pre_boxes = result[0]['boxes'].cpu().detach().numpy()
    pre_labels = result[0]['labels'].cpu().detach().numpy()
    pre_scores = result[0]['scores'].cpu().detach().numpy()
    pre_masks = result[0]['masks'].cpu().detach().numpy()

    selected_indices = pre_scores > self.bbox_thresh
    box = pre_boxes[selected_indices]
    label = pre_labels[selected_indices]
    score = pre_scores[selected_indices]
    mask = pre_masks[selected_indices]

    return box, label, score, mask

---

bbox_xywh, label, score, mask = mask_detector(ori_img)
print(label)
mmm = label == 1
bbox_xywh = bbox_xywh[mmm]
score = score[mmm]
print(score)
print(bbox_xywh)
##############################################################
[1 1 1]
[0.9999887 0.99995875 0.9999497 ]
[[377.16418 72.82647 531.6579 378.49054 ]
[315.75507 107.948395 346.40503 189.95299 ]
[108.50862 66.49912 219.38882 344.93854 ]]