Detecting small objects in large images

[mansi-aggarwal-2504]

I have some images which are 6000 × 4000 pixels in size ( relatively large images) and the objects that have been labelled here are quite small. I trained the model and had to set the --img-size parameter to 1280 as anything higher than that was throwing RuntimeError: CUDA out of memory error.

When I ran detect.py, I received zero detections. Even at 0.1 conf-thres.

Any idea how I can detect very small objects in large images?

[tetelevm]

As far as I know, YOLO has a minimum object size in the image. I would advise you to split the image into 9 images with an overlap of 100 -> (0-40%), (30-70%), (60-100%) in width and height and search for objects on them with a smaller network.

[mansi-aggarwal-2504]

I also had that idea but do you know of any way I could adjust the labels to cropped images. Since I have ground truth labels (xml and txt) for the full images.

[Melhaya]

I also did some cropping to images I am using for training. reverse the normalization of the labels and adjust the X-center and Y-Center according to the cropping you did and then Normalize them back again. It will take a couple of trials to get it correct but it works. Just make sure you visualize the new labels on your cropped images as a sanity check. :)

[glenn-jocher]

@tetelevm @mansi-aggarwal-2504 @Melhaya actually a simpler solution might be to train on native resolution crops and then to run inference at batch-size 1 with native resolution. You can do this using the new Albumentations integration in #3882 and using i.e. A.RandomCrop(width=1280, height=1280).

[mansi-aggarwal-2504]

Hi @glenn-jocher, thank you for this tip.
Just wanted to clarify, when you said "train on native resolution crops", you're still suggesting I crop the 6000x4000 images into smaller images (say 6 smaller images), and run inference on similar cropped images also?

[glenn-jocher]

@mansi-aggarwal-2504 I wouldn't crop anything. It's a large developmental expense to create that pipeline.

[mansi-aggarwal-2504]

@glenn-jocher right. So you're saying I can use the 6000x4000 images and train on them and run the inference at batch-size 1 with same resolution images. And benefit from the new Albumentations integration.
Just making sure I understand you correctly.

[glenn-jocher]

@mansi-aggarwal-2504 yes you can train at full size with random crops, and then run inference at full size.

[mansi-aggarwal-2504]

Thank you for your help @glenn-jocher!

[fcakyon]

@mansi-aggarwal-2504 @Melhaya @tetelevm you can perform sliced inference with lower image sizes and lower gpu memory usage on a yolov5 model using SAHI package: https://github.com/obss/sahi