about regression targets

[wenhe-jia]

In your code，the regression targets of rcnn is （dx，dy，dw，dh），but in your evaluation script，it seems like that you just take out the outputs from bbox_pred_1st/2nd/3rd layers and use them as the four locations （x1，y1，x2，y2）of rois，can you explain this for me？

[zhaoweicai]

You can have a look at the deploy.prototxt. All (dx, dy, dw, dh) have been decoded as (x1, y1, x2, y2) as final output.

[wenhe-jia]

Got it, thank you very much.

[wenhe-jia]

Hi, is me again. In your decode_bbox_layer.cpp, you screen out rois with high iou(>0.95), then you add gt_bboxes as rois in proposal_target_layer, what is this for? Can you explain this for me?

[zhaoweicai]

It is just to remove redundant gt boxes, since gt boxes will be added in proposal_target_layer each time.

[wenhe-jia]

OK, thanks a lot.

In addition, the channel of bbox_pred in rcnn is 8, I think is fg/bg delta for each roi, but not num_classes*4. Why did you do like this, does this bring benefits for bbox regression in your experiments?

[zhaoweicai]

It is class-agnostic bbox regression. In my experiments, I don't find it has difference with class-specific bbox regression.