Semi-supervised Detector Training with Prototypes for Vehicle Detection

Research output: Chapter in Book/Report/Conference proceedingConference contributionResearchpeer-review

Abstract

Abstract— Adapting detectors to new datasets is needed in scenarios where a user has a specific dataset that contains novel classes or is recorded in a setting where a pretrained detector fails. While detectors based on Convolutional Neural Networks (CNNs) are state-of-the-art and nowadays publicly available, they suffer from bad generalization capabilities when applied on datasets that notably differ from the one they were trained on. Finetuning the detector is only possible if the dataset is large enough to not destroy the underlying feature representation. We propose a method where only a few prototypes are labeled for training in a semi-supervised manner. In particular, we separate the detection from the classification step to avoid impairing the bounding box proposal generation. Our trained prototype classification network provides labels to automatically source a large dataset containing 20 to 30 times more samples without further supervision, which we then use to train a more powerful network. We evaluate our method on a private vehicle dataset with six classes and show that evaluating on a previously unseen recording site we can gain an accuracy increase of 9% at same precision and recall levels. We further show that finetuning with as few as 25 labeled samples per class doubles accuracy compared to directly using pretrained features for nearest neighbor classification.
Original languageEnglish
Title of host publication2019 IEEE Intelligent Transportation Systems Conference, ITSC 2019
PublisherInstitute of Electrical and Electronics Engineers
Number of pages6
Publication statusAccepted/In press - Jun 2019
Event22nd IEEE International Conference on Intelligent Transportation Systems - Auckland, New Zealand
Duration: 27 Oct 201930 Oct 2019
https://www.itsc2019.org/

Conference

Conference22nd IEEE International Conference on Intelligent Transportation Systems
Abbreviated titleITSC 2019
CountryNew Zealand
CityAuckland
Period27/10/1930/10/19
Internet address

Cite this

Waltner, G., Opitz, M., Krispel, G., Possegger, H., & Bischof, H. (Accepted/In press). Semi-supervised Detector Training with Prototypes for Vehicle Detection. In 2019 IEEE Intelligent Transportation Systems Conference, ITSC 2019 Institute of Electrical and Electronics Engineers.

Semi-supervised Detector Training with Prototypes for Vehicle Detection. / Waltner, Georg; Opitz, Michael; Krispel, Georg; Possegger, Horst; Bischof, Horst.

2019 IEEE Intelligent Transportation Systems Conference, ITSC 2019. Institute of Electrical and Electronics Engineers, 2019.

Research output: Chapter in Book/Report/Conference proceedingConference contributionResearchpeer-review

Waltner, G, Opitz, M, Krispel, G, Possegger, H & Bischof, H 2019, Semi-supervised Detector Training with Prototypes for Vehicle Detection. in 2019 IEEE Intelligent Transportation Systems Conference, ITSC 2019. Institute of Electrical and Electronics Engineers, 22nd IEEE International Conference on Intelligent Transportation Systems, Auckland, New Zealand, 27/10/19.
Waltner G, Opitz M, Krispel G, Possegger H, Bischof H. Semi-supervised Detector Training with Prototypes for Vehicle Detection. In 2019 IEEE Intelligent Transportation Systems Conference, ITSC 2019. Institute of Electrical and Electronics Engineers. 2019
Waltner, Georg ; Opitz, Michael ; Krispel, Georg ; Possegger, Horst ; Bischof, Horst. / Semi-supervised Detector Training with Prototypes for Vehicle Detection. 2019 IEEE Intelligent Transportation Systems Conference, ITSC 2019. Institute of Electrical and Electronics Engineers, 2019.
@inproceedings{04423cd665254ab29043bbe10c1fca8a,
title = "Semi-supervised Detector Training with Prototypes for Vehicle Detection",
abstract = "Abstract— Adapting detectors to new datasets is needed in scenarios where a user has a specific dataset that contains novel classes or is recorded in a setting where a pretrained detector fails. While detectors based on Convolutional Neural Networks (CNNs) are state-of-the-art and nowadays publicly available, they suffer from bad generalization capabilities when applied on datasets that notably differ from the one they were trained on. Finetuning the detector is only possible if the dataset is large enough to not destroy the underlying feature representation. We propose a method where only a few prototypes are labeled for training in a semi-supervised manner. In particular, we separate the detection from the classification step to avoid impairing the bounding box proposal generation. Our trained prototype classification network provides labels to automatically source a large dataset containing 20 to 30 times more samples without further supervision, which we then use to train a more powerful network. We evaluate our method on a private vehicle dataset with six classes and show that evaluating on a previously unseen recording site we can gain an accuracy increase of 9{\%} at same precision and recall levels. We further show that finetuning with as few as 25 labeled samples per class doubles accuracy compared to directly using pretrained features for nearest neighbor classification.",
author = "Georg Waltner and Michael Opitz and Georg Krispel and Horst Possegger and Horst Bischof",
year = "2019",
month = "6",
language = "English",
booktitle = "2019 IEEE Intelligent Transportation Systems Conference, ITSC 2019",
publisher = "Institute of Electrical and Electronics Engineers",
address = "United States",

}

TY - GEN

T1 - Semi-supervised Detector Training with Prototypes for Vehicle Detection

AU - Waltner, Georg

AU - Opitz, Michael

AU - Krispel, Georg

AU - Possegger, Horst

AU - Bischof, Horst

PY - 2019/6

Y1 - 2019/6

N2 - Abstract— Adapting detectors to new datasets is needed in scenarios where a user has a specific dataset that contains novel classes or is recorded in a setting where a pretrained detector fails. While detectors based on Convolutional Neural Networks (CNNs) are state-of-the-art and nowadays publicly available, they suffer from bad generalization capabilities when applied on datasets that notably differ from the one they were trained on. Finetuning the detector is only possible if the dataset is large enough to not destroy the underlying feature representation. We propose a method where only a few prototypes are labeled for training in a semi-supervised manner. In particular, we separate the detection from the classification step to avoid impairing the bounding box proposal generation. Our trained prototype classification network provides labels to automatically source a large dataset containing 20 to 30 times more samples without further supervision, which we then use to train a more powerful network. We evaluate our method on a private vehicle dataset with six classes and show that evaluating on a previously unseen recording site we can gain an accuracy increase of 9% at same precision and recall levels. We further show that finetuning with as few as 25 labeled samples per class doubles accuracy compared to directly using pretrained features for nearest neighbor classification.

AB - Abstract— Adapting detectors to new datasets is needed in scenarios where a user has a specific dataset that contains novel classes or is recorded in a setting where a pretrained detector fails. While detectors based on Convolutional Neural Networks (CNNs) are state-of-the-art and nowadays publicly available, they suffer from bad generalization capabilities when applied on datasets that notably differ from the one they were trained on. Finetuning the detector is only possible if the dataset is large enough to not destroy the underlying feature representation. We propose a method where only a few prototypes are labeled for training in a semi-supervised manner. In particular, we separate the detection from the classification step to avoid impairing the bounding box proposal generation. Our trained prototype classification network provides labels to automatically source a large dataset containing 20 to 30 times more samples without further supervision, which we then use to train a more powerful network. We evaluate our method on a private vehicle dataset with six classes and show that evaluating on a previously unseen recording site we can gain an accuracy increase of 9% at same precision and recall levels. We further show that finetuning with as few as 25 labeled samples per class doubles accuracy compared to directly using pretrained features for nearest neighbor classification.

UR - https://www.tugraz.at/institute/icg/research/team-bischof/learning-recognition-surveillance/people/waltner/

UR - https://files.icg.tugraz.at/f/d77fa8eb28334e1d9b96/?dl=1

M3 - Conference contribution

BT - 2019 IEEE Intelligent Transportation Systems Conference, ITSC 2019

PB - Institute of Electrical and Electronics Engineers

ER -