Default parameter values

[woctezuma]

Hello,

Thank you for sharing this work as a PyPI package!

I have questions about default parameter values.

retinaface/retinaface/inference.py

Lines 39 to 43 in 29611c7

	arg("-m", "--max_size", type=int, help="Resize the largest side to this number", default=960)
	arg("-b", "--batch_size", type=int, help="batch_size", default=1)
	arg("-j", "--num_workers", type=int, help="num_workers", default=12)
	arg("--confidence_threshold", default=0.7, type=float, help="confidence_threshold")
	arg("--nms_threshold", default=0.4, type=float, help="nms_threshold")

Are these values recommended based on the training dataset and procedure of RetinaFace?
Or are these values which worked best for you?

More specifically, I don't know which value for max_size to choose for my personal dataset of mixed photos and hand-drawn/computer-assisted artworks. I am analysing 300x450 images. So, maybe I should upsample them 2x, and go for 900 max size.

As for the confidence threshold, I would settle for something closer to 0.5 (or even less, to detect hand-drawn faces), and lower the NMS threshold as well (to avoid detecting too many boxes). I understand that it might not be the case for people using the detector on natural photography. Do you think it is a bad idea to lower the threshold?

[woctezuma]

Apparently, training data was 840x840 resolution.

retinaface/retinaface/configs/2020-07-20.yaml

Line 86 in 29611c7

image_size: [840, 840]

[maxlund]

Yet example code uses 960 in both the inference examples, and in the readme we see 2048 instead, even though it is loading the same pretrained model where you can see the 840x840 image size in its config.

model = get_model("resnet50_2020-07-20", max_size=2048)
model.eval()
annotation = model.predict_jsons(image)

I have been using 960 successfully 🤷