Is it normal/better to use photo collages (multiple photos in one image) as a dataset instead of single images for training the object detection SSD model?

I am using Tensorflow Object Detection API to make an SSD model to identify and detect different leave diseases. My dataset consists of 256x256 leaves (plant village dataset), I want my model to be trained for detecting multiple objects as well with different sizes (most of the leaves are almost as big as the image itself - which gives bad accuracy for small objects) so I thought combining those pictures into the collages to include several leaves in each picture and creating a dataset with the following way, for each label I will use single leaf images, 4 leaves images (2x2 collage) and 9 images (3x3 collage). Do you think that it will give better results for multiple object detection or Should stick with single images and it won't change anything? If it works what would you recommend to set for the image resize option? Should it be still low like 300x300 or it will be better to have 800x800 since having 9 images in some part?


1 Answer 1


1.You can use Progressive Resizing in your dataset in order to improve your results. Link here
This method can be helpful when the images only depict parts of the sample.

2.Also, in order for your model to be good at predicting smaller leaves that do not take the entire images, you should make your validation set balanced ie. a number of smaller leave images and leaves that take the whole image should be balanced in the validation set.

3.I don't know much about the collage-based approach, but you can check this

Hope it helps. Please upvote if it helped.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.