KITTI VISUAL ODOMETRY DATASET

README

## http://cvlibs.net/datasets/kitti/eval_semantics.php
## https://omnomnom.vision.rwth-aachen.de/data/rwth_kitti_semantics_dataset.zip

### DATASET FOR SEMANTIC SEGMENTATION

fine_tune.py

IMG_MEAN = np.array((104.00698793,116.66876762,122.67891434), dtype=np.float32)

BATCH_SIZE = 2
DATA_DIRECTORY = './../rwth_kitti_semantics_dataset/'
DATA_LIST_PATH = './../rwth_kitti_semantics_dataset/splits/train_tf.txt'
IGNORE_LABEL = 255
INPUT_SIZE = '321,321'
LEARNING_RATE = 1e-4
NUM_CLASSES = 14
NUM_STEPS = 20000
RANDOM_SEED = 1234
RESTORE_FROM = './deeplab_resnet.ckpt'
SAVE_NUM_IMAGES = 2
SAVE_PRED_EVERY = 10
SNAPSHOT_DIR = './snapshots_finetune/'

label_colours.py

# colour map
label_colours = [(0,   0,   0),
         (0,  0, 255),
         (255,  0,  0),
           (255,255,  0),
           (  0,255,  0),
           (255,  0,255),
           (  0,255,255),
           (255,  0,153),
           (153,  0,255),
           (  0,153,255),
           (153,255,  0),
           (255,153,  0),
           (  0,255,153),
           (  0,153,153)]

masks_kitti.ipynb

• Download dataset from https://omnomnom.vision.rwth-aachen.de/data/rwth_kitti_semantics_dataset.zip
• Convert RGB label images using masks_kitti.ipynb
• Download deeplab_resnet.ckpt
• Change fine_tune.py configuration (example is given)
• Change https://github.com/DrSleep/tensorflow-deeplab-resnet/blob/master/deeplab_resnet/image_reader.py#L115 to decode_png
• Run fine_tune.py

Training results

good job！

what is the purpuse of IMG_MEAN in fine_tune.py