Error while running Example of TrainAnSSDObjectDetector

hello! please help me out to solve this

when i run the example "https://www.mathworks.com/help/vision/ref/trainssdobjectdetector.html#mw_aadddb45-dcc1-4968-8d1d-d9ec600dc23e", i receive following error:

Error using matlab.io.datastore.TransformedDatastore/read (line 162)

Invalid transform function defined on datastore.

Error in nnet.internal.cnn.GeneralDatastoreDispatcher>iReadNonEmptyOneRow (line 644)

data = read(ds);

Error in nnet.internal.cnn.GeneralDatastoreDispatcher (line 224)

exampleDataFromDatastore = iReadNonEmptyOneRow(ds);

Error in nnet.internal.cnn.DataDispatcherFactory.createDataDispatcherMIMO (line 170)

nnet.internal.cnn.GeneralDatastoreDispatcher( ...

Error in vision.internal.cnn.trainNetwork>iCreateTrainingDataDispatcher (line 190)

dispatcher = nnet.internal.cnn.DataDispatcherFactory.createDataDispatcherMIMO( ...

Error in vision.internal.cnn.trainNetwork (line 40)

trainingDispatcher = iCreateTrainingDataDispatcher(ds, mapping, trainedNet,...

Error in trainSSDObjectDetector (line 233)

[network, info] = vision.internal.cnn.trainNetwork(...

Caused by:

Error using tabular/rowfun>dfltErrHandler (line 517)

Applying the function '@(varargin)iSelectTrainingSamples(params,varargin{:})' to the 1st row of A generated the following error:

The value of 'bboxB' is invalid. Expected input to be finite.

[Huma Hafeez]

i am using MATLAB 2020a

Answer 1

[Divya Gaddipati]

From the error "The value of 'bboxB' is invalid. Expected input to be finite.", it seems like one or more of the groundtruth bounding boxes have NaNs/Inf values. Make sure your groundtruths are valid and non-empty i.e., the values of the bounding boxes are finite, positive, non-fractional, non-NaN and should be within the image boundary with a positive height and width. You should either discard or fix the samples with invalid bounding boxes.

[David]

This answer is insufficient. There are obvious reasons for wanting to train on images that have no target in them. My observation is you get a ton of false positives in areas of images that the detector has never seen examples of.

I understand Mathworks has acknowledged this deficiency and is working on a FIX, can you give a better workaround or a date when we can actually make use of the SSD detector. As it is it's useless.