Experimental Study of the Suitability of CNN-based Holistic Descriptors for Accurate Visual Localization
| Type of publication: | Inproceedings |
| Citation: | jaenal2019 |
| Booktitle: | Proceedings of the 2Nd International Conference on Applications of Intelligent Systems |
| Series: | APPIS '19 |
| Year: | 2019 |
| Pages: | 28:1--28:6 |
| Publisher: | ACM |
| Location: | Las Palmas de Gran Canaria, Spain |
| Address: | New York, NY, USA |
| ISBN: | 978-1-4503-6085-2 |
| URL: | http://doi.acm.org/10.1145/330... |
| DOI: | 10.1145/3309772.3309800 |
| Abstract: | Holistic Image Descriptors (HIDs) are compact representations of a whole image that, being suitable for Place Recognition, are not appropriate for accurate Visual Localization. The most successful HIDs are those extracted from Convolutional Neural Networks (CNNs) like VGG, ResNet, InceptionV4 or NetVLAD. Very recently, the equivariance property has been proposed to reflect how image 2D transformations (e.g. rotation, flip, scale changes) influence the descriptor [17]. Our work experimentally analyzes whether such property can be a good indicator of the suitability of the existing CNN-based HID for estimating changes in the camera pose, which produces more complex transformations of the image than the pure transformations analyzed in [17]. The results we report here are a preliminary work in the context of an ongoing project towards appearance-based localization of autonomous mobile robots. |
| Userfields: | img_url=http%3A%2F%2Fmapir.isa.uma.es%2Fimagesrepo%2Fpapers%2F2019%2F2019_jaenal_APPIS.png,rank_indexname=,rank_pos_in_category=,rank_num_in_category=,rank_cat_name=,impact_factor= |
| Keywords: | |
| Authors | |
| Added by: | [] |
| Total mark: | 0 |
|
Attachments
|
|
|
|
|
Notes
|
|
|
|
|
|
Topics
|
|
|
|
|
