There are some interesting literature about RPNs (Region Proposal Network). The most concise and helpful documentation that I found so far is the following: https://www.quora.com/How-does-the-region-proposal-network-RPN-in-Faster-R-CNN-work?share=1.

But there is something that I still don't understand through my various lectures. RPNs are designed to propose several candidate regions. From which, a selection will be done to know which candidates fits our needs.

But, RPNs and neural network in general are deterministic. Thus, once trained, they will always produce the same output for a given input; there is no way to query new candidates given the same input image. As far as I understood, RPNs are trained to produce a fix number of proposal, for each new image. But how the training work then? If the RPN has to produce 300 candidates, what should be the labeled data that we use for the training, knowing that a training image probably won't have more than 5 golden truth bounding boxes?

And then, knowing that the bounding box sizes are not consistent among candidates, how does the CNN behind operates with the different size of the input?


The first answer in your commented link answers one point about how region proposals are selected. It is the Intersection Over Union (more formally the Jaccard Index) metric. So how much of your anchor overlaps the label. There is usually a lower limit set for this metric to then filter out all the useless proposals, and the remaining matches can be sorted, choosing the best.

I recommend reading through this excellently explained version of a proposal network - Mask-R-CNN (Masked Region-based CNN). If you prefer looking at code, there is the full repo here, implemented in Keras/Tensorflow (there is also a PyTorch implementation linked somewhere).

There is even an explanatory Jupyter notebook, which might help make things click for you.

  • $\begingroup$ Indeed. But my question was not so much about the selection of bounding box, but more about the training. I mean by that - whenever we do a backpropagation, we need to compute the loss function, and have every output compared to the golden value. But since we could have 300x4 output, and we have for instance 5x4 truth (golden) output, how do we do for the backpropagation and the training? $\endgroup$
    – Emile D.
    Jun 21 '18 at 14:33

To know how RPN work for training, we can dive into the code wrote by Matterport, which is 10,000 stared and tf/keras implementation Mask R-CNN repo.

You can check the build_rpn_targets function in mrcnn/model.py

If we used the generated anchors (depends on your anchor scales, ratio, image size ...) to calculate the IOU of anchors and ground truth,

    # Compute overlaps [num_anchors, num_gt_boxes]
    overlaps = utils.compute_overlaps(anchors, gt_boxes)

we can know how overlaps between anchors and ground truth. Then we choose positive anchors and negative anchors based on their IOU with ground truth. According to Mask R-CNN paper, IOU > 0.7 will be positive anchors and < 0.3 will be negative anchors, otherwise will be neutral anchors and not used when training

    # 1. Set negative anchors first. They get overwritten below if a GT box is
    # matched to them. 
    anchor_iou_argmax = np.argmax(overlaps, axis=1)
    anchor_iou_max = overlaps[np.arange(overlaps.shape[0]), anchor_iou_argmax]
    rpn_match[anchor_iou_max < 0.3] = -1
    # 2. Set an anchor for each GT box (regardless of IoU value).
    # If multiple anchors have the same IoU match all of them
    gt_iou_argmax = np.argwhere(overlaps == np.max(overlaps, axis=0))[:,0]
    rpn_match[gt_iou_argmax] = 1
    # 3. Set anchors with high overlap as positive.
    rpn_match[anchor_iou_max >= 0.7] = 1

To effectively train RPN, you need to set up the RPN_TRAIN_ANCHORS_PER_IMAGE carefully to balance training if there is few objects in one image. Please note that there can be multiple anchors match one ground truth since we can give the bbox off-set for each anchor to fit the ground truth.

Hope the answer is clear for you!

  • $\begingroup$ Well, it is indeed with more information. But the question I had is: inference time, we still propose 300 regions (or whatever RPN_TRAIN_ANCHORS_PER_IMAGE is). How do we select the few good ones from the proposed regions? And from what I gather, we backpropagate by checking the error with the IOU. But we could be missing a region not proposed during the forward pass. Thus, we could converge to have all the propositions at the same location? $\endgroup$
    – Emile D.
    Mar 25 '19 at 20:02
  • $\begingroup$ In inference time, RPN simply sends all positive ROIs (predict as foreground) to the second stage. Since most ROIs may overlap with each other, we'll apply Non Maximum Suppression to keep the highest probability ROIs. The main purpose of RPN is trying to catch all foreground object in an image. So RPN tends to make some false positive, which can be fixed on the second stage classifier. $\endgroup$
    – jimmy15923
    Apr 1 '19 at 4:35
  • $\begingroup$ Thank you for the clarification for the inference time, it is clearer to me. But in training time then, all the bounding boxes will eventually converge to the same proposals? Is there some criterion to avoid too much similarities between the proposals to avoid convergence to few same locations? $\endgroup$
    – Emile D.
    Apr 9 '19 at 22:46
  • $\begingroup$ As I mentioned, Non-Maximum Suppression will be applied to all anchor boxes during training and inferencing, so most proposals will not overlap too much with other proposals. The threshold of NMS is a hyperparameter which defines in config.py. $\endgroup$
    – jimmy15923
    Apr 10 '19 at 7:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.