Error using vision.int​ernal.rcnn​.BBoxTrain​ingDataDis​patcher/ge​t.NumObser​vations (line 191) when training my own dataset

Hi guys, is there anyone who used the trainRCNNObjectDetector to his/her own dataset? I did this but encountered the problem of "Error using vision.internal.rcnn.BBoxTrainingDataDispatcher/get.NumObservations (line 191)", followed by " Too many outputs requested. Most likely cause is missing [] around left hand side that has a comma separated list expansion."

I used the same code to train the stopSign images as in example from https://www.mathworks.com/help/vision/examples/object-detection-using-deep-learning.html, there is no such problem, but when I changed the stopSigns ground truth table to the truth table that I created myself, there comes up this problem.

[Vivian Zhao]

To be added, the former two stages are going successfully. Network training complete. The problem comes in Step 3 of 3: Training bounding box regression models for each object class...

[Mir Pias]

Did you solve the problem? same thing happens to me when I change the data

Answer 1

[Michael Doxey]

Hi Vivian,

I understand that you are having issues when trying to use the trainRCNNObjectDetector function when trying to train your own dataset. It seems when you are trying to train your own data, you are encountering an error telling you that you have requested "too many outputs."

Unfortunately, this is a currently known bug. However, there is a current workaround. The workaround is to make your ground truth boxes larger than 32x32, by resizing the images and scaling the boxes before training. After resizing, the trainRCNNObjectDetector function should work on your own datasets.

I hope this workaround was able to help accomplish what you were hoping to.

[Massimo Del Guasta]

Hi I've the same problem as Vivian, my training images are all 310*310*1 pixels (monochrome). I don't understand how to avoid the problem. Any help?