Sliced Inference Support for Improved Results

[mali-afridi]

This pull request supports sliced window inference to improve object detection results, especially for larger or high-resolution images. By slicing an image into smaller overlapping windows and running the detection model on each slice, we increase the chances of detecting smaller objects or objects partially visible in a single frame.
The key addition is a slicing mechanism that divides an image into smaller sections before passing it through the detection model. After detection, the predictions from each slice are merged back into the original image space. This method helps detect objects that span slice boundaries while also handling scenarios where a large object might be missed in a single inference pass on the full image.

Key Changes:

1. The infer.py script now has a slice window inference option, which is enabled by the -sliced flag.
2. Introduced functions to slice images into smaller tiles with a configurable overlap ratio.
3. Implemented postprocessing logic to merge detection results from individual slices back into a single set of predictions.
4. Support for setting the number of slices via the --numberofboxes flag.

How to Use:

1. When running the inference script, use the—- sliced flag to enable the sliced window inference.
2. You can configure the number of slices by specifying the—- number of boxes argument, which defines how many slices to create based on the image's aspect ratio.

Results on 0.6 threshold score:

These results were achieved by using

Model	Dataset	Input Size	APval	AP50val	#Params(M)	FPS	checkpoint
rtdetr_r101vd	COCO	640	54.3	72.8	76	74	url*

With Sliced Inference	Without Sliced Inference

Example Usage

python tools/infer.py \ -c configs/rtdetr/rtdetr_r101vd_6x_coco.yml \ -r /path/to/model/weights.pth \ --im-file='/path/to/image.jpg' \ --device=cuda:0 \ -s True \ -nc 35

[lyuwenyu]

Can you add some message for this feature in Updates in README.md?

[mali-afridi]

Done.