TAD16K: An enhanced benchmark for autonomous driving (Source Code)

Here is the source code of our paper TAD16K: an enhanced benchmark for autonomous driving. (http://autopilot.qq.com/ICIP2017/ or backup link: http://url.cn/53k4BPv, password: mXo0Af)

We provide the evaluation results of two state-of-the-art object detection algorithms (SSD and DetectNet) on our benchmark TAD16K, which can be used as the baseline for future comparison purpose.

SSD can be obtained at https://github.com/weiliu89/caffe/tree/ssd.

DetectNet can be obtained at https://github.com/NVIDIA/DIGITS/tree/master/examples/object-detection.

We bind Caffe as a module in Torch7 and maintain an updated torch-caffe-binding source code at: https://github.com/lymhust/caffe_torch_binding.

You have to have installed and built Caffe, then do this:

CAFFE_DIR=/*path-to-caffe-root*/ luarocks make

Then SSD and DetectNet can be easily evaluated under Torch7 environment.

How to evaluate:

1. Download TAD16K, pretrained models and results at http://autopilot.qq.com/ICIP2017/.
2. Replace the empty folders (models_DetectNet, models_SSD and results).
3. Run test_totxt.lua to generate evaluation results of DetectNet and SSD.
4. Run drawPRCurve_all.lua to generate PR curve for each method.
5. Run drawPRCurve_twoMethods.lua to generate PR curve of two methods in one coordinate.

• Before running, please goto 'utils/evaluate' and build the source:

mkdir build
cd ./build
cmake ..
make

Useful functions:

• test_basler.lua: test two methods using Basler industrial camera.
• test_imgs.lua: test two methods using images stored in one folder.
• test_webcamera.lua: test two methods using web camera.
• detectnet.lua: parse the caffe model of DetectNet.
• ssd.lua: parse the caffe model of SSD.
• nms.lua: non maximum suppression.
• json2txt.lua: parse the json file provided by TAD16K and save it as txt format.
• data_statistics.lua: statistics of the four types of objects in TAD16K.

Examples:

require 'camera'
-- Types:  light, vehicle, pedestrian and sign
pretype = 'sign'
-- Methods: DetectNet and SSD
local method = 'DetectNet'

if (method == 'DetectNet') then
	dofile('detectnet.lua')
elseif (method == 'SSD') then
	dofile('ssd.lua')
else
	print('Unknow method!')
end

-- Testing using web camera
camera = image.Camera{idx=0, width=im_w, height=im_h, fps=30}

while (true) do
  local start = sys.clock()
  frame = camera:forward()
  frame = image.scale(frame, im_w, im_h, 'simple')
  frame, result = process_one(frame[{{},{topend,bottomend},{}}])
  win = image.display{win=win, image=frame}
  local time = sys.clock() - start
  print("FPS: ".. 1/time)
  print("Time: "..(time*1000)..'ms\n')
end

DetectNet results on PointGrey frames: