Problem with segmentation my dataset in DeepLab

I adapted the "build_convert" of the PASCAL Dataset to convert my dataset to execute the DeepLab, but all my segmentation tests generate a black image prediction.

My dataset has images with 140x140 px. (see an example above:)

Original Image

Original Mask

Converted Mask

Prediction Image (complete black, all prediction pixels are background)

What am I doing wrong?
Could anybody help me with this problem?

My step-by-step:

1. Convert the segmentation masks

python ./remove_gt_colormap.py
--original_gt_folder="${SEG_FOLDER}"
--output_dir="${SEMANTIC_SEG_FOLDER}"
SEG_FOLDER -> the original mask images

2. Create the tfrecord files
   (I reused the PASCAL example)

python ./build_voc2012_data.py
--image_folder="${IMAGE_FOLDER}"
--semantic_segmentation_folder="${SEMANTIC_SEG_FOLDER}"
--list_folder="${LIST_FOLDER}"
--output_dir="${OUTPUT_DIR}"

3. To train

NUM_ITERATIONS=100
python "${WORK_DIR}"/train.py
--logtostderr
--train_split="val"
--model_variant="${modelo}"
--atrous_rates=12 --atrous_rates=24 --atrous_rates=36 --output_stride=8
--decoder_output_stride=4
--train_crop_size="513,513"
--train_batch_size=4
--training_number_of_steps="${NUM_ITERATIONS}"
--train_logdir="${TRAIN_LOGDIR}"
--dataset_dir="${PASCAL_DATASET}"

4. to eval
   python "${WORK_DIR}"/vis.py
   --logtostderr
   --vis_split="val"
   --model_variant="${modelo}"
   --atrous_rates=12 --atrous_rates=24 --atrous_rates=36 --output_stride=8 --decoder_output_stride=4
   --vis_crop_size="513,513"
   --checkpoint_dir="${TRAIN_LOGDIR}"
   --vis_logdir="${VIS_LOGDIR}"
   --dataset_dir="${PASCAL_DATASET}"
   --max_number_of_iterations=1

Thanks

Thanks