predict_kitti.lua

#! /usr/bin/env luajit

--[[

This script computes the 3 pixel error on all KITTI 2012 training examples with
the fast architecture. Alternatively, if the `action` variable is set to
'submit', the disparity maps for all test image pairs are computed and stored
so that they can be submitted to the KITTI evaluation server. You should see a
2.81% 3 pixel out-noc error rate on the server.

Don't use this script to fit hyperparameters; the error is computed on the
training examples.

This is not the fastest way to use the neural network---a new process is
spawned and the network is loaded from disk for each image pair---but
is probably the safest.

Usage
-----

   $ ./predict_kitti.lua
   0   0.0028267929719645
   1   0.026568045683624
   2   0.039333925127797
   ...
   191 0.078452818068974
   192 0.012351983422143
   193 0.066736774940625
   0.03222369495401

]]--

require 'image'
require 'torch'
require 'libadcensus'

action = 'test'
assert(action == 'test' or action == 'submit')

path = 'data.kitti/unzip'
cmd = './main.lua kitti fast -a predict' ..
   ' -net_fname net/net_kitti_fast_-a_train_all.t7' ..
   ' -left %s -right %s -disp_max 228'

if action == 'test' then
   err_sum = 0
   n_te = 194
   dir = 'training'
elseif action == 'submit' then
   n_te = 195
   dir = 'testing'
end

for i = 0, n_te - 1 do
   -- call mc-cnn
   local im0 = ('%s/%s/image_0/%06d_10.png'):format(path, dir, i)
   local im1 = ('%s/%s/image_1/%06d_10.png'):format(path, dir, i)
   local im = image.loadPNG(im0)
   local img_height = im:size(2)
   local img_width = im:size(3)
   os.execute(cmd:format(im0, im1) .. ' > /dev/null')
   local disp = torch.FloatTensor(torch.FloatStorage('disp.bin')):view(1, 1, img_height, img_width)

   if action == 'test' then
      -- ground truth
      local ground_truth = torch.FloatTensor(1, img_height, img_width)
      adcensus.readPNG16(ground_truth, ('%s/training/disp_noc/%06d_10.png'):format(path, i))

      -- compute the error
      local mask = torch.ne(ground_truth, 0):float()
      local bad = torch.add(disp, -1, ground_truth):abs():gt(3):float():cmul(mask)
      local err = bad:sum() / mask:sum()
      err_sum = err_sum + err
      print(i, err)
   elseif action == 'submit' then
      adcensus.writePNG16(disp, img_height, img_width, ("out/%06d_10.png"):format(i))
      print(i)
   end

   collectgarbage()
end

if action == 'test' then
   print(err_sum / n_te)
end