![](bg6.png)
to 90% of recall when classifier are used by the sched-
uler but other tradeoff have be evaluated to ensure that
this tradeoff selection does not bias the evaluation.
However, even with a grid search on all tradeoff, all
baselines are always at least 30% lower in F
1
measure
than the scheduler.
Thus, this experiment confirms the relevancy of
such scheduler at least for time constraint batch clas-
sification.
Currently, the very high F
1
score of the scheduler
could not be expected from the calibration (see table
1). As, all classifiers are calibrated on validation set
to have a common recall of 90%, precision should be
the only mutable value. This is however not the case
due to interaction in the cascaded decision: our sys-
tem has a lower recall but achieves higher precision
than expected. We believe that this could be explained
by the difference between calibration which is done
classifier per classifier and the scheduling were boxes
are finally classified by multiples classifiers. In other
words, the considered classifiers have some comple-
mentarity that are freely exploited by the scheduler.
Anyway, this does not invalidate the evaluation. A
more careful study of this last result is out of the scope
of this paper whose main result (in our opinion) is that
the scheduler is able to produce sufficiently good out-
put.
5 CONCLUSION
The goal of this paper is to make a step toward sched-
ulers able to help the integration of large computer
vision library into complex robotic system. To make
a first step, we chose a scheduling problem derived
of one the simplest computer vision: time constraint
batch classification. We describe a scheduling frame-
work for this problem. We apply it on car detec-
tion on real remote sensing images with realistic set-
tings in term of time constrains, target hardware (hy-
brid CPU-GPU) and computer vision classifiers (with
deep learning ones). The results of this experiment
is that our scheduler goes further than expected by
achieving state of the art results - while it is just de-
signed to produce a sufficiently good results to be rel-
evant for prototyping/integrating requirements.
REFERENCES
Bellman, R. (1957). A markovian decision process. Journal
of Mathematics and Mechanics.
Dalal, N. and Triggs, B. (2005). Histograms of oriented
gradients for human detection. In CVPR.
Ferrero, R., Rivera, J., and Shahidehpour, S. (1998). A
dynamic programming two-stage algorithm for long-
term hydrothermal scheduling of multireservoir sys-
tems. Power Systems.
Girshick, R., Donahue, J., Darrell, T., and Malik, J. (2014).
Rich feature hierarchies for accurate object detection
and semantic segmentation. In CVPR.
Graham, R. L., Lawler, E. L., Lenstra, J. K., and Kan, A. R.
(1979). Optimization and approximation in determin-
istic sequencing and scheduling: a survey. Annals of
discrete mathematics.
Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J.,
Girshick, R., Guadarrama, S., and Darrell, T. (2014).
Caffe: Convolutional architecture for fast feature em-
bedding. In ICM.
Lagrange, A., Le Saux, B., Beaupere, A., Boulch, A., Chan-
Hon-Tong, A., Herbin, S., Randrianarivo, H., and
Ferecatu, M. (2015). Benchmarking classification of
earth-observation data: from learning explicit features
to convolutional networks. In IGARSS.
Lamiraux, F. and Laumond, J. (2000). Smooth path plan-
ning for car-like robots. In ISAS.
Lazebnik, S., Schmid, C., and Ponce, J. (2006). Beyond
bags of features: Spatial pyramid matching for recog-
nizing natural scene categories. In CVPR.
Lucas, B. D., Kanade, T., et al. (1981). An iterative image
registration technique with an application to stereo vi-
sion. In IJCAI.
Marti, K. and Qu, S. (1998). Path planning for robots by
stochastic optimization methods. Journal of Intelli-
gent and robotic Systems.
Metta, G., Fitzpatrick, P., and Natale, L. (2006). Yarp: yet
another robot platform. IJARS.
Ouelhadj, D. and Petrovic, S. (2009). A survey of dy-
namic scheduling in manufacturing systems. Journal
of Scheduling.
Quigley, M., Conley, K., Gerkey, B. P., Faust, J., Foote, T.,
Leibs, J., Wheeler, R., and Ng, A. Y. (2009). Ros: an
open-source robot operating system. In ICRA.
Russakovsky, O., Li, L.-J., and Fei-Fei, L. (2015). Best of
both worlds: human-machine collaboration for object
annotation. In CVPR.
Shan, Q., Jia, J., and Agarwala, A. (2008). High-quality mo-
tion deblurring from a single image. In ACM Trans-
actions on Graphics.
Shotton, J., Winn, J., Rother, C., and Criminisi, A. (2006).
Textonboost: Joint appearance, shape and context
modeling for multi-class object recognition and seg-
mentation. In ECCV.
Song, H., Liu, C.-C., Lawarr
´
ee, J., and Dahlgren, R. W.
(2000). Optimal electricity supply bidding by markov
decision process. Power Systems.
Trapeznikov, K. and Saligrama, V. (2013). Supervised se-
quential classification under budget constraints. In
ICAIS.
Van Den Berg, J., Abbeel, P., and Goldberg, K. (2011). Lqg-
mp: Optimized path planning for robots with motion
uncertainty and imperfect state information. IJRR.
Zhang, Y. J. (1996). A survey on evaluation methods for
image segmentation. Pattern recognition.
VISAPP 2017 - International Conference on Computer Vision Theory and Applications
352