Are the pretrained nets expecting RGB or BGR inputs?

[josh-wende]

Looking at the code, it is non-obvious to me. Appreciate the help.

[liyufan]

The network expects BGR inputs during inference, you can use the following code to verify in Jupyter Notebook:

import cv2
import numpy as np
import torch
from PIL import Image

from clrnet.datasets import build_dataloader
from clrnet.utils.config import Config

cfg = Config.fromfile("configs/clrnet/clr_resnet18_tusimple.py")
cfg.gpus = 1
cfg.seed = 0

test_loader = build_dataloader(cfg.dataset.test, cfg, is_train=False)
test_set = test_loader.dataset

idx = 2
# (C, H, W) -> (H, W, C)
Image.fromarray((test_set[idx]["img"].permute(1, 2, 0).numpy() * 255).astype(np.uint8))

In the second cell:

img = cv2.imread(test_set[idx]["meta"].data["full_img_path"])[cfg.cut_height :, :, :]
Image.fromarray(cv2.resize(img, (cfg.img_w, cfg.img_h)))

The two images are the same. Because cv2.imread reads the images in BGR order, the input to the network is also BGR.