Unsupervised Multi-Target Domain Adaptation for Object Detection

FPV@IPLAB - Department of Mathematics and Computer Science, University of Catania, Italy

Giovanni Pasqualino, Antonino Furnari, Giovanni Maria Farinella


Domain adaptation approaches are aimed to efficiently train object detectors by leveraging labeled synthetic images, inexpensively generated from a 3D model, and unlabeled real images, which are cheaper to obtain than labeled ones. Most of the state-of-the-art techniques consider only one source and one target domain for the adaptation task. However, real world scenarios often involve many target domains which can easily arise, for instance, from the use of different cameras during the acquisition of the images. In this work, we investigate whether the availability of multiple unlabeled target domains can improve domain adaptive object detection algorithms. To study the problem, we propose a new dataset comprising images of 16 different objects rendered from a 3D model and collected in a real environment using two different cameras. We experimentally assess that current domain adaptive object detectors can improve their performance by leveraging the multiple targets and introduce a new unsupervised multi-target domain adaptation approach for object detection which outperform current methods.

Dataset



We propose a dataset of synthetic and real images related to 16 artworks present in "Galleria regionale Palazzo Bellomo" located in Siracusa, Italy. The dataset contains two set of images, synthetic and real which are divided has follows:

Synthetic Dataset

  • Training set: 51284 images
  • Validation set: 24525 images
  • Test set: 23960 images
  • Real Hololens Dataset

  • Training set: 1502 images
  • Test set: 688 images
  • Real GoPro Dataset

  • Training set: 1911 images
  • Test set: 796 images

  • You can download the whole dataset and annotations at this link

    Methods



    We explore the following methods:
    1) baseline approaches without adaption;
    2) domain adaptation through image to image translation;
    3) domain adaptation through feature alignment;
    4) new multi target domain adaptation method MDA-RetinaNet (see the figure below);
    5) domain adaptation combining feature alignment and image to image translation.



    MDA-RetinaNet architecture.

    Code



    The MDA-RetinaNet architecture code is available at this link

    Results



    Qualitative Results

    Paper

    G. Pasqualino, A. Furnari, G. M. Farinella, "Unsupervised Multi-Target Domain Adaptation for Object Detection", Submitted to International Conference on Image Processing 2021.

    Acknowledgement

    This research has been supported by the project VALUE (N. 08CT6209090207 - CUP G69J18001060007) - PO FESR 2014/2020 - Azione 1.1.5., by Research Program Pia.ce.ri. 2020/2022 Linea 2 - University of Catania, and by MIUR AIM - Linea 1 - AIM1893589 - CUP E64118002540007.