Batch inference

[iu110]

A great job, are there any tips on setting up bounding boxes to perform batch inferencing?

[limuloo]

@iu110 Thank you for your recognition of our work. Do you mean to include images with varying numbers of boxes in a single batch, or should the number of boxes be the same in all images within a batch?

[iu110]

Thank you for getting back to me. Your current code only provides inference for a single image. If I want to infer multiple different images simultaneously, such as the 800 samples you provided, do I need to set up multiple sets of bounding boxes manually? Would you happen to have any tips for this? After all, it may be difficult to set them manually.

[limuloo]

@iu110 If you want to infer multiple different images simultaneously, you can do it this way:

Taking COCO-MIG as an example, where each image contains at most 6 instances, you can pad those images with fewer than 6 instances using empty text and a [0, 0, 0, 0] bounding box to reach 6 instances, which will have minimal impact on the results. Once all images are ensured to contain 6 instances, you can then place different images into one batch for inference.

[mengtan00]

@iu110 If you want to infer multiple different images simultaneously, you can do it this way:

Taking COCO-MIG as an example, where each image contains at most 6 instances, you can pad those images with fewer than 6 instances using empty text and a [0, 0, 0, 0] bounding box to reach 6 instances, which will have minimal impact on the results. Once all images are ensured to contain 6 instances, you can then place different images into one batch for inference.

Hi, thanks for your great work. I tried to do the batch inference following the above way, but found the MIGCProcessor only supports batch_size=1. How can I modify the code to make it work and will you release the code that supports batch inference?

[limuloo]

@mengtan00 Thank you for your attention to our work. Unfortunately, I am too busy with work recently and do not have time to implement it myself😭. You can try to modify MIGC_processor to achieve the above functions.