Keras CV Usage on Custom dataset for Object detection

AI Discussions
,
1

Hi Everyone,
I keenly need to access Keras CV Object detection model for my research work,
Link to docs - Object Detection with KerasCV
I am not getting the desired output when training on custom dataset

Someone please help

2

Hi @Atharv_Gulati

Is the linked document your personal project? can I know what kind of desired output you are looking for?

Regards
DP

3

Hi Deepti,
Thanks for replying,
So this is the documentation of Keras CV to train the model on custom dataset.
I want to do that for my research but when running the code in that docs as it is, it is only showing true boxes not predicted ones

Hense I need help to train KerasCV on custom dataset

4

hi @Atharv_Gulati

can I know if you have completed DLS and tensorflow advanced technique specialisation here?

Also can I know if you are using different custom dataset then mentioned on the website?

Also how are using the detection boxes, I mean how are you assigning the true box and predicted box, remember the steps in the link of keras.io shares yolo8 object detection codes and then uses the pretrained model and recalling the prediction_decoder where you threshold is redjusted as well as confidence threshold.

iou_threshold controls the threshold of intersection over union (IoU) that two boxes must have in order for one to be pruned out.

I hope you are applying this logic to your custom dataset.

Rather than sharing the link to the kers.io website, share where you have run down the codes with the results you have got, so one can point you in the right direction.

Also in case you have completed Tensorflow advanced technique specialisation, I would recommend you to check the graded assignment related to the bounding box.

I am adding @ai_curious who is also someone experienced in this topic, so he can also put forth his valuable suggestions.

@ai_curious thank you in advance.

Regards
DP

5

I only did a quick read of the linked page. The only places I see a ‘true’ bounding box visualization is the very first image of the cat and in the advanced section where it suggests visualization of ground truth bounding boxes to confirm the eval_ds has been loaded correctly. The bounding boxes visualized in the middle section, the dog and potted plants, and at the end of the advanced section, look to me to be predicted boxes (y_pred). Not clear to me what the OP means above only showing true boxes

Also, the only reason to go through the effort of training on a custom dataset is if the pre trained model performs poorly - bad localization precision or misclassification. This can only be determined empirically- not visually.

1 Like
6

Hi @ai_curious
I completely understand what you are saying here but if you check prediction on custom dataset, the model shows only blue boxes which are the true labels as specified in y_true
and no yellow boxes of y_pred
I have used the code similarly as shown, even the dataset is the same

7

Unfortunately, the visualization package of KerasCV seems to be missing from both their documentation web site and their github repository so it is rather difficult to assist. Suggest try interacting with the developers directly at the github repo.