transfer learn without losing old memoy YoloX

Hi,

I have learned to perform transfer learning on yoloX object detector....However my new data is only based on "person" class (with person climbing a wall).

[imds,blds] = objectDetectorTrainingData(gTruth);
classes = categories(blds.LabelData{1,2});

>>classes 1x1 cell array {'person'}

detector=yoloxObjectDetector("small-coco",classes,"InputSize",targetDataSize);
detector = trainYOLOXObjectDetector(trainingData,detector,options);

When i train trainYOLOXObjectDetector() on this new data, it fits the model to detect only new data type but "person" images recognized by original yoloX model are no longer detected.... How can i perform transfer learning in this scenario.

When i keep classes equal to yoloX 80 classes, and train model on new data, it says class sizes are mismatch.

I don't want to train on mix data data (coco images + new data) as it will be tedious...

How can i pass new data while maintaining detection capability of original yoloX...

Thanks

Answer 1

[Matt J]

Continual learning is a serious research topic.There are no off-the-shelf tools in Matlab to do it in a sophisticated way. You would probably have to use a lower level training routine like trainnet or possibly a custom training loop. This paper might also be relevant:

https://dl.acm.org/doi/epdf/10.1145/3641584.3641651

[tren]

Ok this means that best i can do is to find an appropriate dataset and and mix my data with it and train, rite....?

my images contain small sized humans detected in very low resolution and vegetation by cctv... Basically person climbing walls/area (captured from 50-100m by cctv camera)

can someone suggest any good dataset,please?

thanks

[Matt J]

You obviously have at least a few examples of the old data type because you've reported to us that they aren't correctly processed with the new yoloX, right? One thing you can try is to make replicas of the few that you have and intersperse them throughout your new data set. If you make enough replicas so that every minibatch contains a few of the old types during training, each gradient step will be hindered from forgetting at least those examples. It might be enough of a regularizing force that the small data set you have will be enough.

[tren]

is there a provision of downloading train dataset through Matlab... I am looking to download images with cctv small human detection type to match my data///

[Matt J]

From ChatGPT:

---------------------------

MATLAB provides helper functions for downloading some standard computer vision datasets (for example COCO, VOC, Caltech Pedestrian in some examples, etc.), but there is not a built-in MATLAB command that automatically downloads a specialized CCTV small-human dataset matched to your application.

For CCTV-style small-person detection, people commonly use datasets such as:

• CrowdHuman
• WiderPerson
• VisDrone
• TinyPerson
• Caltech Pedestrian Dataset

These are usually downloaded manually from the dataset website, then imported into MATLAB using:

imageDatastore

boxLabelDatastore

combine

If your targets are very small (typical CCTV wide-angle scenes), datasets like TinyPerson or VisDrone are often more representative than generic COCO pedestrians.

MATLAB examples sometimes include helper scripts such as:

helper.downloadDataset(...)

but those helpers are example-specific and not a general dataset repository service.

Also note that for YOLOX or YOLOv8-style training, matching:

• camera height,
• viewing angle,
• compression artifacts,
• night/day conditions,
• and target pixel size

often matters more than sheer dataset size. Generic pedestrian datasets can perform poorly on tiny CCTV humans unless fine-tuned on domain-matched imagery