cuteboyqq

you can use YoloV4 or YoloV5 mosaic class (__getitem__ function will output all mosaic images and corresponding labels) to generate mosaic data, but you need to translate labels to label.txt...

l run the code of process_cityscapes.py put at firectory dataset/cityscape_scripts cd dataset/cityscape_scripts python process_cityscapes.py

command add --use_lovaszsoftmax for example : python train.py --model ESPNet_v2 --dataset cityscape --batch_size 2 --use_lovaszsoftmax

l also have same issue when l train the model of FPENet and DABNet and ESPNet, the highest miou is about 53% when training

l suggust you do not do multi scale aug and rand crop, becuase your dataset just focus on the drivable area, the rand crop will crop the sky only, and...

sorry, you also need to modified code if you set MULTI_SCALE false below is base_dataset.py modified code ``` # ------------------------------------------------------------------------------ # Modified based on https://github.com/HRNet/HRNet-Semantic-Segmentation # ------------------------------------------------------------------------------ import cv2 import...