hoijanlai / fcn8-free-space-detection

Semantic Segmentation

info

trained with tensorflow 1.3 python version: 3.5 platform: ubuntu 16.04

usage

to start a train

python main.py <batch_size> <epochs> <learning_rate> <keep_prob> <regularization_factor>

(check main.py to see default parameters)

some params worth trying

batch_size=2, epochs=30, keep_prob=0.5, learning_rate=0.0001, regularization=0.01

architecture

FCN 8, please check this paper kernel size was 4x4 for the first-stage upsampling, 8x8 for the second and 32x32 for the final(8 times).

details

• augmentations : rotation, flip, zoom, shift and channel shift. tf.contrib.keras.preprocessing.image check get_batch_function in helper.py NOTE: Instead of using keras's implementation, channel_shift is implemented by shearing the channel axis, which actually just shuffles along the channel axis.

• super high channel-shift probability in the most recent run:

  get_batches_fn = helper.gen_batch_function(os.path.join(data_dir, 'data_road/training'),
                                             image_shape,
                                             aug_size=0.6,
                                             channel_shift=0.8)

  since I found the model pretty confused on shaded road. It seems to work compared with the ones without channel shift, but need more epochs to train better (the not-so-shaded examples is well segmented with 20 epochs, but the shaded ones need extra 40 epochs to get a acceptable result).

• in main.py, set freeze argument to True to only train the skip layers:

def optimize(nn_last_layer, correct_label, learning_rate, num_classes, reg=1e-2, freeze=False):

the most recent trial

batch_size=2, epochs=60, keep_prob=0.5, learning_rate=0.0001, regularization=0.01