multi_scale parameter used to resize_square instead of affine transformation about yolov3 HOT 3 CLOSED

Ah, --multi-scale randomizes the image size from 320 to 608 pixels, in steps of 32, like darknet. The affine transformation has a scale parameter that zooms in or out the image contents, but it does not change the image size in pixels. Just like your iphone can zoom in 1X or 2X or 10X, and still output the same 8MP image size, the affine scale parameter is analogous to this zoom.

I suppose for training it is accomplishing a similar effect, except that if the affine transform scales an image by 2X for example, then most of it will disappear off the edges, and only the centermost area will remain, whereas --multi-scale from 320 to 608 will result in the same image contents, just at a higher resolution.

from yolov3.

As @glenn-jocher said - just scaling (but keeping image size constant) will result in cropping the image at random locations. The multi-scale training actually resizes images, which achieves large invariance to object sizes, as well as (and this is the real kicker with YOLOv3) a single model for 3 different scales (320, 416, 608).

from yolov3.

@nirbenz absolutely, as you highlighted, multi-scale training with different image sizes (320, 416, 608) aids in acquiring robustness to object sizes and allows for a single model to represent multiple scales. This is advantageous for YOLOv3. The approach ensures effective handling of objects at various scales while maximizing the model's versatility across different scenarios. All credit goes to the YOLO community for fostering such innovative techniques.

from yolov3.