0

I am training a custom dataset in Tensorflow 1.15.3 on top of mobilenet_v3_large_seg with standard research models. The training in general is ok but the borders/edges are very noisy, what might help to improve detecting segments with better boundaries?

python deeplab/train.py --logtostderr --training_number_of_steps=84000 \
--output_stride 4 \
--train_split="trainval" --model_variant="mobilenet_v3_large_seg" \
--train_crop_size="384,384" --dataset="123" --save_interval_secs=300 \
--save_summaries_secs=300 --save_summaries_images=True --log_steps=100 --train_logdir=123/training \
--dataset_dir=123 --min_scale_factor=0.75 --max_scale_factor=1.25 \
--fine_tune_batch_norm=True --train_batch_size=4 --scale_factor_step_size=0.05
Thomas
  • 1,622
  • 17
  • 24

1 Answers1

1

With further testing and tweaking two thing improved my results:

  1. Noisy images are challenging to segmentation, so I smooth the images before with mean/median filter for instance.

  2. Training really needs lot of samples and epochs and will improve slowly.

Thomas
  • 1,622
  • 17
  • 24