无边界框训练数据及NYU RGBD数据集上Mask-RCNN训练方法咨询

Great question—training Mask-RCNN without explicit bounding box annotations is totally feasible, especially since you only need to validate the instance segmentation functionality. Let’s start with your specific NYU RGBD dataset scenario, then dive into the general approach for unboxed training data.

Since NYU RGBD provides dense instance masks (even without pre-defined bounding boxes), we can leverage these masks to generate pseudo bounding boxes and adapt the standard Mask-RCNN pipeline:

• Generate pseudo bounding boxes from instance masks
  The simplest and most effective way is to compute the minimum enclosing rectangle for each instance mask. This gives you a valid bounding box that aligns perfectly with the mask. You can implement this with OpenCV:

  import cv2
  import numpy as np
  
  def mask_to_pseudo_bbox(mask):
      # Mask is a binary numpy array (1 = instance, 0 = background)
      contours, _ = cv2.findContours(mask.astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
      if contours:
          x, y, w, h = cv2.boundingRect(contours[0])
          return [x, y, x + w, y + h]  # Convert to (x1, y1, x2, y2) format
      return None
  

  Iterate over all your instance masks to generate corresponding pseudo boxes, then format your dataset into the standard Mask-RCNN input structure (each sample has image, bboxes, class labels, and masks).

• Leverage NYU RGBD's depth channel for better performance
  Unlike standard RGB datasets, NYU RGBD includes depth information. You can modify Mask-RCNN's backbone to accept 4-channel input (RGB + Depth) instead of 3-channel RGB. This extra spatial information will help the model better distinguish instances, especially in cluttered indoor scenes.

• Sync data augmentation for masks and pseudo boxes
  When applying data augmentations (rotation, scaling, cropping), make sure to adjust both the masks and pseudo boxes simultaneously to avoid misalignment. Libraries like Albumentations handle this automatically if you configure the bbox_params correctly:

  import albumentations as A
  from albumentations.pytorch import ToTensorV2
  
  train_transform = A.Compose([
      A.RandomRotate90(p=0.5),
      A.Resize(height=512, width=512),
      A.Normalize(mean=[0.485, 0.456, 0.406, 0.5], std=[0.229, 0.224, 0.225, 0.25]),
      ToTensorV2()
  ], bbox_params=A.BboxParams(format='pascal_voc', label_fields=['class_labels']))
  

For any dataset without bounding box annotations (but with instance-level masks), these general strategies work:

• Pseudo bounding box generation (core approach)
  Beyond minimum enclosing rectangles, you can use rotated bounding boxes for irregular instances, or even mask centroid-based boxes if needed. If you only have semantic segmentation maps (not instance masks), first convert them to instance-level masks using methods like watershed segmentation or a pre-trained instance segmentation model (e.g., Mask2Former) to generate pseudo instance masks, then derive boxes from those.

• Adjust loss weighting to prioritize mask segmentation
  If your pseudo boxes are not perfectly accurate, you can reduce the weight of the bounding box regression loss in Mask-RCNN's total loss function. For example, set bbox_loss_weight = 0.1 instead of the default 1.0 to let the model focus more on mask quality while still learning to refine boxes.

• Weakly/semi-supervised training (for limited data)
  If you have a small dataset, start with a pre-trained Mask-RCNN (e.g., COCO weights) and fine-tune it on your pseudo box data. Then, use the model's own predictions as pseudo labels for a second round of training—this self-supervised loop can help the model refine its mask and box predictions without additional manual annotations.

• Use Mask-RCNN variants designed for box-free training
  While standard Mask-RCNN works with pseudo boxes, some variants like MaskScoring R-CNN or SOLOv2 are optimized for instance segmentation without relying heavily on bounding boxes. These models can be adapted to your dataset with minimal changes if you want to avoid pseudo box generation entirely.

内容的提问来源于stack exchange，提问作者nagaharish