dlib's MMOD+CNN detector with 7 conv layers [MMOD+CNN]

Submitted on 22 Nov. 2017 02:34 by
Davis King (dlib)

Running time:0.28 s
Environment:4 cores @ >3.5 Ghz (C/C++)

Method Description:
The submission essentially shows what happens when you run dlib's
MMOD+CNN vehicle detection example program (i.e. on the kitti
cars object detection benchmark. The method uses the max-margin
object detection loss with a simple CNN with 7 conv layers, batch
normalization, and relu units. The network runs fully
convolutionally over an image pyramid to find the cars. This
sliding window method produces detection boxes which are then
refined by a random forest regression to give boxes tightly fit
around each car.
See the .tar file with the code for the full training
details. The file contains everything needed to
reproduce the results.
Latex Bibtex:

Detailed Results

Object detection and orientation estimation results. Results for object detection are given in terms of average precision (AP) and results for joint object detection and orientation estimation are provided in terms of average orientation similarity (AOS).

Benchmark Easy Moderate Hard
Car (Detection) 89.86 % 83.14 % 69.29 %
Car (Orientation) 36.43 % 33.90 % 28.49 %
This table as LaTeX

2D object detection results.
This figure as: png eps pdf txt gnuplot

Orientation estimation results.
This figure as: png eps pdf txt gnuplot

eXTReMe Tracker