Cross Training for Pedestrian recognition using Convolutional Neural networks

Danut Ovidiu Pop (1, 2, 3) , Alexandrina Rogozan (2) , Fawzi Nashashibi (1) , Abdelaziz Bensrhair (2)

1 RITS - Robotics & Intelligent Transportation Systems
2 LITIS - Laboratoire d'Informatique, de Traitement de l'Information et des Systèmes
3 UBB - Babes-Bolyai University [Cluj-Napoca]

In recent years, deep learning classification methods, specially Convolutional Neural Networks (CNNs), combined with multi-modality image fusion schemes have achieved remarkable performance. Hence, in this paper, we focus on improving the late-fusion scheme for pedestrian classification on the Daimler stereo vision data set. We propose cross training method in which a CNN for each independent modality (Intensity, Depth, Flow) is trained and validated on different modalities, in contrast to classical training method in which the training and validation of each CNN is on same modality. The CNN outputs are then fused by a Multi-layer Perceptron (MLP) before making the recognition decision.

Mots clés: convolutional neural network; pedestrian recognition; deep learning; late; fusion; cross; training; late-fusion; cross-training

Domaines: Vision par ordinateur et reconnaissance de formes [cs.CV]

Danut Ovidiu Pop, Alexandrina Rogozan, Fawzi Nashashibi, Abdelaziz Bensrhair. Cross Training for Pedestrian recognition using Convolutional Neural networks. ORASIS 2017, GREYC, Jun 2017, Colleville-sur-Mer, France. ⟨hal-01866658⟩