FCN-32s Fully Convolutional Semantic Segmentation on NYUDv2

readme.md

name	caffemodel	caffemodel_url	sha1	gist_id
FCN-32s Fully Convolutional Semantic Segmentation on NYUDv2	fcn-32s-nyud.caffemodel	http://dl.caffe.berkeleyvision.org/fcn-32s-nyud.caffemodel	ee9bb63261d30160dfb4675b564b912d335eeaf5	16db1e4ad3afc2614067

This is a model from the paper:

Fully Convolutional Networks for Semantic Segmentation
Jonathan Long, Evan Shelhamer, Trevor Darrell
arXiv:1411.4038

This is the single stream, 32 pixel prediction stride version, trained on the NYUDv2 RGBD dataset.

This model was trained for the 40-class segmentation task defined by Gupta et al. (CVPR13). The final layer outputs scores for each class, which may be normalized via softmax or argmaxed to obtain per-pixel labels.

Two inputs are expected.

1. The visible channels in BGR order, with the following per-channel mean subtracted:

   B 104.00698793 G 116.66876762 R 122.67891434

2. The HHA-encoded depth, as defined by Gupta et al. (ECCV14) and available from https://github.com/s-gupta/rcnn-depth, with the following per-channel mean subtracted:

   H 132.431 H 94.076 A 118.477

This is a pre-release: it requires unmerged PRs to run. It should be usable with the branch available at https://github.com/longjon/caffe/tree/future. Training ought to be possible with that code, but the original training scripts have not yet been ported.

This model is reported in the paper as FCN-32s RGB-HHA. It obtains 32.8 mean IU and 48.0 frequency weighted mean IU on the NYUDv2 test set.

fcn-32s-nyud-deploy.prototxt

name: 'FCN'

input: 'bgr_data'
input_dim: 1
input_dim: 3
input_dim: 500
input_dim: 500

input: 'hha_data'
input_dim: 1
input_dim: 3
input_dim: 500
input_dim: 500

# bgr network
layers { bottom: 'bgr_data' top: 'conv1_1' name: 'conv1_1' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 64 pad: 100 kernel_size: 3 } }
layers { bottom: 'conv1_1' top: 'conv1_1' name: 'relu1_1' type: RELU }
layers { bottom: 'conv1_1' top: 'conv1_2' name: 'conv1_2' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 64 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv1_2' top: 'conv1_2' name: 'relu1_2' type: RELU }
layers { name: 'pool1' bottom: 'conv1_2' top: 'pool1' type: POOLING
  pooling_param { pool: MAX kernel_size: 2 stride: 2 } }
layers { name: 'conv2_1' bottom: 'pool1' top: 'conv2_1' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 128 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv2_1' top: 'conv2_1' name: 'relu2_1' type: RELU }
layers { bottom: 'conv2_1' top: 'conv2_2' name: 'conv2_2' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 128 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv2_2' top: 'conv2_2' name: 'relu2_2' type: RELU }
layers { bottom: 'conv2_2' top: 'pool2' name: 'pool2' type: POOLING
  pooling_param { pool: MAX kernel_size: 2 stride: 2 } }
layers { bottom: 'pool2' top: 'conv3_1' name: 'conv3_1' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 256 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv3_1' top: 'conv3_1' name: 'relu3_1' type: RELU }
layers { bottom: 'conv3_1' top: 'conv3_2' name: 'conv3_2' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 256 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv3_2' top: 'conv3_2' name: 'relu3_2' type: RELU }
layers { bottom: 'conv3_2' top: 'conv3_3' name: 'conv3_3' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 256 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv3_3' top: 'conv3_3' name: 'relu3_3' type: RELU }
layers { bottom: 'conv3_3' top: 'pool3' name: 'pool3' type: POOLING
  pooling_param { pool: MAX kernel_size: 2 stride: 2 } }
layers { bottom: 'pool3' top: 'conv4_1' name: 'conv4_1' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 512 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv4_1' top: 'conv4_1' name: 'relu4_1' type: RELU }
layers { bottom: 'conv4_1' top: 'conv4_2' name: 'conv4_2' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 512 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv4_2' top: 'conv4_2' name: 'relu4_2' type: RELU }
layers { bottom: 'conv4_2' top: 'conv4_3' name: 'conv4_3' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 512 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv4_3' top: 'conv4_3' name: 'relu4_3' type: RELU }
layers { bottom: 'conv4_3' top: 'pool4' name: 'pool4' type: POOLING
  pooling_param { pool: MAX kernel_size: 2 stride: 2 } }
layers { bottom: 'pool4' top: 'conv5_1' name: 'conv5_1' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 512 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv5_1' top: 'conv5_1' name: 'relu5_1' type: RELU }
layers { bottom: 'conv5_1' top: 'conv5_2' name: 'conv5_2' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 512 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv5_2' top: 'conv5_2' name: 'relu5_2' type: RELU }
layers { bottom: 'conv5_2' top: 'conv5_3' name: 'conv5_3' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 512 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv5_3' top: 'conv5_3' name: 'relu5_3' type: RELU }
layers { bottom: 'conv5_3' top: 'pool5' name: 'pool5' type: POOLING
  pooling_param { pool: MAX kernel_size: 2 stride: 2 } }
layers { bottom: 'pool5' top: 'fc6' name: 'fc6' type: CONVOLUTION
  convolution_param { engine: CAFFE kernel_size: 7 num_output: 4096 } }
layers { bottom: 'fc6' top: 'fc6' name: 'relu6' type: RELU }
layers { bottom: 'fc6' top: 'fc6' name: 'drop6' type: DROPOUT
  dropout_param { dropout_ratio: 0.5 } }
layers { bottom: 'fc6' top: 'fc7' name: 'fc7' type: CONVOLUTION
  convolution_param { engine: CAFFE kernel_size: 1 num_output: 4096 } }
layers { bottom: 'fc7' top: 'fc7' name: 'relu7' type: RELU }
layers { bottom: 'fc7' top: 'fc7' name: 'drop7' type: DROPOUT
  dropout_param { dropout_ratio: 0.5 } }
layers { name: 'score' type: CONVOLUTION bottom: 'fc7' top: 'score'
  convolution_param { engine: CAFFE num_output: 40 kernel_size: 1 } }

# hha network
layers { bottom: 'hha_data' top: 'conv1_1_hha' name: 'conv1_1_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 64 pad: 100 kernel_size: 3 } }
layers { bottom: 'conv1_1_hha' top: 'conv1_1_hha' name: 'relu1_1_hha' type: RELU }
layers { bottom: 'conv1_1_hha' top: 'conv1_2_hha' name: 'conv1_2_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 64 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv1_2_hha' top: 'conv1_2_hha' name: 'relu1_2_hha' type: RELU }
layers { name: 'pool1_hha' bottom: 'conv1_2_hha' top: 'pool1_hha' type: POOLING
  pooling_param { pool: MAX kernel_size: 2 stride: 2 } }
layers { name: 'conv2_1_hha' bottom: 'pool1_hha' top: 'conv2_1_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 128 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv2_1_hha' top: 'conv2_1_hha' name: 'relu2_1_hha' type: RELU }
layers { bottom: 'conv2_1_hha' top: 'conv2_2_hha' name: 'conv2_2_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 128 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv2_2_hha' top: 'conv2_2_hha' name: 'relu2_2_hha' type: RELU }
layers { bottom: 'conv2_2_hha' top: 'pool2_hha' name: 'pool2_hha' type: POOLING
  pooling_param { pool: MAX kernel_size: 2 stride: 2 } }
layers { bottom: 'pool2_hha' top: 'conv3_1_hha' name: 'conv3_1_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 256 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv3_1_hha' top: 'conv3_1_hha' name: 'relu3_1_hha' type: RELU }
layers { bottom: 'conv3_1_hha' top: 'conv3_2_hha' name: 'conv3_2_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 256 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv3_2_hha' top: 'conv3_2_hha' name: 'relu3_2_hha' type: RELU }
layers { bottom: 'conv3_2_hha' top: 'conv3_3_hha' name: 'conv3_3_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 256 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv3_3_hha' top: 'conv3_3_hha' name: 'relu3_3_hha' type: RELU }
layers { bottom: 'conv3_3_hha' top: 'pool3_hha' name: 'pool3_hha' type: POOLING
  pooling_param { pool: MAX kernel_size: 2 stride: 2 } }
layers { bottom: 'pool3_hha' top: 'conv4_1_hha' name: 'conv4_1_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 512 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv4_1_hha' top: 'conv4_1_hha' name: 'relu4_1_hha' type: RELU }
layers { bottom: 'conv4_1_hha' top: 'conv4_2_hha' name: 'conv4_2_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 512 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv4_2_hha' top: 'conv4_2_hha' name: 'relu4_2_hha' type: RELU }
layers { bottom: 'conv4_2_hha' top: 'conv4_3_hha' name: 'conv4_3_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 512 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv4_3_hha' top: 'conv4_3_hha' name: 'relu4_3_hha' type: RELU }
layers { bottom: 'conv4_3_hha' top: 'pool4_hha' name: 'pool4_hha' type: POOLING
  pooling_param { pool: MAX kernel_size: 2 stride: 2 } }
layers { bottom: 'pool4_hha' top: 'conv5_1_hha' name: 'conv5_1_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 512 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv5_1_hha' top: 'conv5_1_hha' name: 'relu5_1_hha' type: RELU }
layers { bottom: 'conv5_1_hha' top: 'conv5_2_hha' name: 'conv5_2_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 512 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv5_2_hha' top: 'conv5_2_hha' name: 'relu5_2_hha' type: RELU }
layers { bottom: 'conv5_2_hha' top: 'conv5_3_hha' name: 'conv5_3_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE num_output: 512 pad: 1 kernel_size: 3 } }
layers { bottom: 'conv5_3_hha' top: 'conv5_3_hha' name: 'relu5_3_hha' type: RELU }
layers { bottom: 'conv5_3_hha' top: 'pool5_hha' name: 'pool5_hha' type: POOLING
  pooling_param { pool: MAX kernel_size: 2 stride: 2 } }
layers { bottom: 'pool5_hha' top: 'fc6_hha' name: 'fc6_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE kernel_size: 7 num_output: 4096 } }
layers { bottom: 'fc6_hha' top: 'fc6_hha' name: 'relu6_hha' type: RELU }
layers { bottom: 'fc6_hha' top: 'fc6_hha' name: 'drop6_hha' type: DROPOUT
  dropout_param { dropout_ratio: 0.5 } }
layers { bottom: 'fc6_hha' top: 'fc7_hha' name: 'fc7_hha' type: CONVOLUTION
  convolution_param { engine: CAFFE kernel_size: 1 num_output: 4096 } }
layers { bottom: 'fc7_hha' top: 'fc7_hha' name: 'relu7_hha' type: RELU }
layers { bottom: 'fc7_hha' top: 'fc7_hha' name: 'drop7_hha' type: DROPOUT
  dropout_param { dropout_ratio: 0.5 } }
layers { name: 'score_hha' type: CONVOLUTION bottom: 'fc7_hha' top: 'score_hha'
  convolution_param { engine: CAFFE num_output: 40 kernel_size: 1 } }

# fuse
layers { type: ELTWISE name: 'fuse' bottom: 'score' bottom: 'score_hha' top: 'score-fused'
  eltwise_param { operation: SUM coeff: 0.5 coeff: 0.5 } }

layers { type: DECONVOLUTION name: 'upsample-new'
  bottom: 'score-fused' top: 'bigscore'
  convolution_param { num_output: 40 kernel_size: 64 stride: 32 bias_term: false } }
layers { type: CROP name: 'crop' bottom: 'bigscore' bottom: 'bgr_data' top: 'upscore' }

Hello Jon,
currently I am reproducing your results on the NYUD dataset and I am asking myself, why you choose 500x500 as input dimensions? The raw images have something around 640*480, so you even must have interpolated the height?!
I'd appreciate any details on the preprocessing!
Thank you ver much!

@timmeinhardt no, actually we never rescale the image dimensions. The 500 x 500 here is merely for sake of example. We reshape the net on-the-fly for each input to process each image at its original dimensions. Sorry there isn't a preprocessing example at the moment. If I have a chance to post a notebook for it I'll follow-up here.