Place recognition for mobile robots is a long-studied topic  due to the far-reaching impact it will have in enabling fully-autonomous systems in the near future. State-of-the-art methods for place-recognition today use hand-engineered image feature descriptors and matching techniques to implement their vision-based loop-closure mechanisms. While these model-based algorithms have enabled significant advances in mobile robot navigation, they are still limited in their ability to learn from new experiences and adapt accordingly. We envision robots to be able to learn from their previous experiences and continuously tune their internal model representations in order to achieve improved task-performance and model efficiency. With these considerations in mind, we introduce a bootstrapped mechanism to learn and fine-tune the model performance of vision-based loop-closure recognition systems in mobile robots.
With a growing set of experiences that a robot logs today, we recognize the need for fully automatic solutions for experience-based task learning and model refinement. Inspired by NetVLAD , we cast the problem of place recognition in mobile robots as a self-supervised metric learning problem. Most previous works [28, 36, 23]
use hand-engineered image descriptors or pre-trained Convolutional Neural Network architectures[19, 39] to describe an image for classification or matching. All these methods, in some way or the other, require a hand-engineered metric
for matching the visual descriptors extracted. The choice of feature extraction needs to be tightly coupled with the right distance metric in order to retrieve similar objects appropriately. This adds yet another level of complexity in designing and tuning reliable systems that are fault tolerant and robust to operating in varying appearance regimes. Furthermore, these approaches do not provide a mechanism to optimize for specific appearance regimes (e.g. learn to ignore fog/rain in those specific conditions). We envision that the distance metric for these feature descriptors be learned[3, 4] from experience, and that it should be done in a bootstrapped manner . Furthermore, we would prefer that the features describing the same place to be repeatably embedded close to each other in some high-dimensional space, with the distances between them to be well-calibrated. With this self-supervised mechanism, we expect robots to be able to quickly adapt to the visual appearance regimes it typically sees, and reliably perform visual place recognition as it gathers more experience.
Ii Related Work
Visual place recognition in the context of vision-based navigation is a well studied problem. In order to identify previously visited locations the system need to be able to extract salient cues from an image that describes the content contained within it. Typically, the same place may be significantly different from its previous appearance due to factors such as variation in lighting (e.g. sunny, cloudy, rainy etc), observed viewpoint (e.g. viewing from opposite directions, viewing from significantly different vantage points), or even perceptual aliasing (e.g. facing and seeing a brick-wall elsewhere). These properties make it challenging to hand-engineer solutions that robustly operate in a wide range of scenarios.
Local and Global methods Some of the earliest forms of visual place recognition entailed directly observing pixel intensities in the image and measuring their correlation. In order to be invariant to viewpoint changes, subsequent works [18, 27, 7, 34, 8, 6] proposed using low-level local and invariant
feature descriptors. These descriptions are aggregated into a single high-dimensional feature vector via Bag-of-Visual-Words (BoVW), VLAD  or Fisher Vectors . Other works [35, 28, 34] directly modeled whole-image statistics and hand-engineered global descriptors such as GIST  to determine an appropriate feature representation for an image.
Sequence-based, Time-based or Context-based methods While image-level feature descriptions are convenient in matching, it becomes less reliable due to perceptual aliasing, or low saliency in images. These concerns led further work [29, 9, 24, 26] in matching whole sequences of consecutive images that effectively describes a place. In SeqSLAM, the authors  identify potential loop closures by casting it as a sequence alignment problem, while others  rely on temporal consistency checks across long image sequences in order to robustly propose loop closures. Mei et al.  finds cliques in the pose graph to extract place descriptions. Lynen et al.  proposed a placeless-place recognition scheme where they match features on the level of individual descriptors. By identifying high-density regions in the distance matrix computed from feature descriptions extracted across a large sequence of images, the system could propose swaths of potentially matching places.
Learning-based methods In one of the earliest works in learning-based methods Kuipers and Beeson  proposed a mechanism to identify distinctive features in a location relative to those in other nearby locations. In FABMAP 
, the authors approximate the joint probability distribution over the bag-of-visual-words data via the Chow-Liu tree decomposition to develop an information-theoretic measure for place-recognition. Through this model, one can sample from the conditional distribution of visual word occurrence, in order to appropriately weight the likelihood of having seen identical visual words before. This reduces the overall rate of false positives, thereby significantly increasing precision of the system. Another work fromLatif et al.  re-cast place-recognition as a sparse convex minimization problem with efficient homotopy methods that enable robust loop-closure hypothesis. In similar light, experience-based learning methods [6, 8] take advantage of the robot’s previous experiences to learn the set of features to match, incrementally adding more details to the description of a place if an existing description is insufficient to match a known place.
Deep Learning methods Recently, the advancements in Convolutional Neural Network (CNN) Architectures [19, 33] have drastically changed the landscape of algorithms used in vision-based recognition tasks. Their adoption in vision-based place recognition for robots [36, 4] have recently shown promising results. However, most domain-specific tasks require further model fine-tuning of these large-scale networks in order to perform reliably well. Despite the ready availability of training models and weights, we foresee the data collection and its supervision being a predominant source of friction for fine-tuning models for tasks such as place recognition in robots. Due to the rich amount of cross-modal information that robots typically collect, we expect to self-supervise tasks such as place recognition by fine-tuning existing CNN models with the experience they have accumulated. To this end, we fine-tune these feature representations specifically for the task of loop-closure recognition and show significant improvements in the precision-recall performance.
|St. Lucia Dataset||Translation ()||Rotation ()||Rot. & Trans. ()||Positive Labels||Negative Labels|
Iii Background: Metric Learning
In this work we rely on metric learning to learn an appropriate metric for the task of place recognition in mobile robots. The problem of metric learning was first introduced as Mahalanobis metric learning in , and subsequently explored  with various dimensionality-reduction, information-theoretic and geometric lenses. More abstractly, metric learning seeks to learn a non-linear mapping that takes in input data pairs , where the Euclidean distance in the new target space is an approximate measure of semantic distance in the original space. We define sets of similar and dissimilar paired examples , and respectively as follows
and define an appropriate loss function that captures the aforementioned properties.
Contrastive Loss The contrastive loss introduced by Chopra et al.  optimizes the distances between positive pairs such that they are drawn closer to each other, while preserving the distances between negative pairs at or above a fixed margin from each other. Intuitively, the overall loss is expressed as the sum of two terms with being the indicator variable in identifying positive examples from negative ones,
Training with Siamese Networks Learning is then typically performed with a Siamese architecture [2, 5], consisting of two parallel networks that share weights amongst each other. The contrastive loss is then defined between the two parallel networks and given by Equation 3. The scalar output loss computed from batches of similar and dissimilar samples are then used to update the parameters of the Siamese network
via Stochastic Gradient Descent (SGD). Typically, batches of positive and negative samples are provided in alternating fashion during training.
Iv Self-Supervised Metric Learning
|Epoch 0||Epoch 30||Epoch 180||
Target GPS Similarity
(From Figure 2)
Iv-a Self-supervised dataset generation
Multi-camera systems and navigation modules have more-or-less become ubiquitous in modern autonomous systems today. Typical systems log this sensory information in an asynchronous manner, providing a treasure of cross-modal information that can be readily used for transfer learning purposes. Here, we focus on the task of vision-based place recognition via a forward-looking camera, by leveraging synchronized information collected via standard navigation modules (GPS/IMU, INS etc.).
Sensor Synchronization In order to formalize the notation used in the following sections, we shall refer to as the synchronized tuple of camera image , and GPS measurement , captured at approximately the same time . In typical systems however, these sensor measurements are captured in an asynchronous manner, and the synchronization needs to be carried out carefully in order to ensure clean and reliable measurements for the bootstrapping procedure. It is important to note that for the specific task of place recognition, can be also be sourced from external sensors such as inertial-navigation systems (INS), or even recovered from a GPS-aided SLAM solution.
Keyframe Sub-sampling While we could consider the full set of synchronized image-GPS pairs, it may be sufficient to learn only from a diverse set of viewpoints. We expect that learning from this strictly smaller, yet diverse set, can substantially speed up the training process while being able to achieve the same performance characteristics when trained with the original dataset. While it is unclear what this sampling function may look like for image descriptions, we can easily provide this measure to determine a diverse set of GPS measurements. We incorporate this via a standard keyframe-selection strategy where the poses are sampled from a continuous stream whenever the relative pose has exceeded a certain translational or rotational threshold from its previously established keyframe. We set these translational and rotational thresholds to 5m, and radians respectively to allow for efficient sampling of diverse keyframes.
Keyframe Similarity The self-supervision is enabled by defining a viewing frustum that applies to both the navigation-view and the image-view . We define a Gaussian similarity kernel between two instances of GPS measurements and given by , (or in short):
where is the GPS translation measured in metric-coordinates at time , and is the corresponding rotation or bearing determined from the sequential GPS coordinates for the particular session (See Figure 2). Here, the only hyper-parameter required is the choice of the bandwidth parameters and , and generally depends on the viewing frustum of the camera used. The resulting self-similarity matrix for the translation (using GPS translation only), and the rotation (using established bearing only) on a single session from the St. Lucia Dataset  is illustrated in Figure 2.
Distance-Weighted Sampling With keyframe based sampling considerably reducing the dataset for efficient training, we now focus on sampling positive and negative pairs in order to ensure speedy convergence of the objective function. We first consider the keyframe self-similarity matrix between all pairs of keyframes for a given dataset, and sample positive pairs whose similarity exceeds a specified threshold . Similarly, we sample negative pairs whose similarity is below . For each of the positive and negative sets, we further sample uniformly by their inverse distance in the original feature space following , to encourage faster convergence.
Iv-B Learning an appropriate distance metric for localization
Our proposed self-supervised place recognition architecture is realized with a Siamese network with an appropriate contrastive loss  (given by Equation 7). This simultaneously finds a reduced dimensional metric space where the relative distances between features in the embedded space are well-calibrated. Here, well-calibrated refers to the property that negative samples are separated at least by a known margin , while positive samples are likely to be separated by a distance less than the margin. Following the terminology in Section III, we consider tuples of similar (positive) and dissimilar (negative) examples for learning an appropriate embedding function . Intuitively, we seek to find a “semantic measure” of distance given by in a target space of such that they respect the kernel defined over the space of GPS measurements as given in Equation 6.
Let be the input data and be the indicator variable representing dissimilar () and similar () pairs of examples within . We seek to find a function that maps the input image to an embedding whose distances between similar places are low, while the distances between dissimilar places are high. We take advantage of availability of synchronized Image-GPS measurements to provide an indicator for place similarity, thereby rendering this procedure fully automatic or self-supervised. Re-writing equation 3 for our problem, we get Equation 7 where measures the “semantic distance” between images (Equation 8).
For brevity, we omit and use instead of the full expression . We pick the thresholds for based on a combination of factors including convergence rate and overall accuracy of the final learned metric. Nominal values of range from 0.8 to 0.9 that indicate the tightness of the overlap between viewing frustums of positive examples, with for negative examples set to 0.4.
Figure 3 illustrates the visual self-similarity matrix of the feature embedding at various stages during the training process on the St. Lucia Dataset (100909_1210). At Epoch 0, when the feature embedding is equivalent to the original feature description, it is hard to disambiguate potential loop-closures due to the uncalibrated nature of the distances. As training progresses, the positively labeled examples of loop-closure image pairs are drawn closer together in the embedded space, while the negative examples are pushed farther from each other. As the training converges, we start to notice a few characteristics in the learned embedding that make it especially powerful in identifying loop-closures: (i) The red diagonal bands in the visual self-similarity matrix are well-separated from the blue background indicating that the learned embedding has identified a more separable function for the purposes of loop-closure recognition; and (ii) The visual self-similarity matrix starts to resemble the target self-similarity matrix computed using the GPS measurements (as shown in Figure 3). Furthermore, the t-SNE embedding 
(colorization) of the learned features extracted at identical locations are strikingly similar, indicating that the learned feature embeddinghas implicitly learned a metric space that is more appropriate for the task of place-recognition in mobile robots.
Iv-C Efficient scene indexing, retrieval and matching
One of the critical requirements for place-recognition is to ensure high recall in loop-closure proposals while maintaining sufficiently high precision in candidate matches. This however requires probabilistic interpretability of the matches proposed, with accurate measures of confidence in order to incorporate these measurements into the back-end pose graph optimization. Similarities or distances measured in the image descriptor space are not well-calibrated, making these measures only amenable to distance-agnostic matching such as -nearest neighbor search. Moreover, an indexing and matching scheme such as -nn also makes it difficult to filter out false positives as the distances between descriptors in the original embedding space is practically meaningless. Calibrating distances by learning a new embedding has the added advantage of avoiding these false positives, while being able to recover confidence measures for the matches retrieved.
Once feature embedding is learned, and the features are mapped to an appropriate target space, we require a mechanism to insert and query these embedded descriptors from a database. We use a KD-Tree in order to incrementally insert features into a balanced tree data structure, thereby enabling queries.
V Experiments and Results
We evaluate the performance of the proposed self-supervised localization method on the KITTI  and St. Lucia Dataset . We compare our approach against the image descriptions obtained from extracting the activations from several layers in the Places365-AlexNet pre-trained model  (conv3, conv4, conv5, pool5, fc6, fc7 and fc8 layers). While we take advantage of the pre-trained models developed in  for the following experiments, the proposed framework could allow us to learn relevant task-specific embeddings from any real-valued image-based feature descriptor. The implementation details of our proposed method is described in detail in section V-C.
V-a Learned feature embedding characteristics
While pre-trained models can be especially powerful image descriptors, they are typically trained on publicly-available datasets such as the ImageNet. that have strikingly different natural image statistics. Moreover, some of these models are trained for the purpose of image or place classification. As with most pre-trained models, we expect some of the descriptive performance of Convolutional Neural Networks to generalize, especially in its lower-level layers (conv1, conv2, conv3). However, the descriptive capabilities in its mid-level and higher-level layers (pool4, pool5, fc layers) start to specialize to the specific data regime and recognition task it is trained for. This has been addressed quite extensively in the literature, arguing the need for domain adaptation and fine-tuning these models on more representative datasets to improve task-specific performance [14, 10].
Similar to previous domain adaptation works [14, 10, 12], we are interested in adapting existing models to newer task domains such as place-recognition with minimal human supervision involved. We argue for a self-supervised approach to model fine-tuning, and emphasize the need for a well-calibrated embedding space, where the features embedded in the new space can provide measures for both similarity and the corresponding confidence associated in matching.
Comparing performance between the original and learned embedding space In Figure 5, we compare the precision-recall performance in loop-closure recognition using the original and learned feature embedding space. For various thresholds of localization accuracy (20 and 30 meters), our learned embedding shows considerable performance boost over the pre-trained Places365-AlexNet model. In the figures, we also illustrate the noticeable drop in performance with the descriptive capabilities in the higher-level layers (fc6, fc7, fc8) as compared to the lower-level layers (conv3, conv4, conv5) in the Places365-AlexNet model. This is as expected, since the higher layers in the CNN (pool5, fc6, fc7) are more tailed to the original classification task they were trained for.
|Precision-Recall comparison at 20m||Precision-Recall comparison at 30m|
Embedding distance calibration As described earlier, our approach to learning an appropriate similarity metric for visual loop-closure recognition affords a probabilistic interpretation of the matches proposed. These accurate measures of confidence can be later used to incorporate these measurements into the back-end pose graph optimization. Figure 8 illustrates the interpretability of the proposed learned embedding distance compared to the original feature embedding distance. The histograms for the embedding distance separation is illustrated for both positive (in green) and negative (in blue) pairs of features. Here, a positive pair refers to feature descriptions of images taken at identical locations, while the negative pairs refer to pairs of feature descriptions that were taken from at least 50 meters apart from each other. The figure clearly illustrates how the learned embedding (Ours-fc7
) is able to tease apart positive pairs, from those between the negative pairs of features, enabling an improved classifier (with a more obvious separator) for place-recognition. Intuitively, the histogram overlap between the positive and negative probability masses measures the ambiguity in loop-closure identification, with our learned feature embedding (Ours-fc7) demonstrating the least amount of overlap.
Nearest-Neighbor search in the learned feature embedding space Once the distances are calibrated in the feature embedding space, even a naïve fixed-radius nearest neighbor strategy, that we shall refer to as -NN, can be surprisingly powerful. In Figure 6, we show that our approach is able to achieve high-recall, with considerably strong precision performance for features that lie within distance (contrastive loss margin as described in Section IV-B) from each other.
|Precision-Recall||Recall for increasing distance|
Furthermore, the feature embedding can also be used in the context of image retrieval with strong recall performance via naïve-Nearest Neighbor (-NN) search. Figure 7 compares the precision-recall performance of the -NN strategy on the original and learned embedding space, and shows a considerable performance gain in the learned embedding space. Furthermore, the recall performance also tends to be higher for the learned embedding space as compared to the original descriptor space.
|Precision-Recall (Ours-fc7)||Recall with increasing k-NN (Ours-fc7)|
|Precision-Recall (Places365-fc7)||Recall with increasing k-NN (Places365-fc7)|
|Separation distance histogram (Ours-fc7)||Separation distance histogram (conv3)||Separation distance histogram (conv4)||Separation distance histogram (conv5)|
|Separation distance histogram (pool5)||Separation distance histogram (fc6)||Separation distance histogram (fc7)||Separation distance histogram (fc8)|
V-B Localization performance within visual-SLAM front-ends
Figure 9 shows the trajectory of the optimized pose-graph leveraging the constraints proposed by our learned loop-closure proposal method. The visual place-recognition module determines constraints between temporally distant nodes in the pose-graph that are likely to be associated with the same physical location. To evaluate the localization module independently, we simulate drift in the odometry chains by injecting noise in the individual ground truth odometry measurements.
|KITTI Sequence 00|
|St. Lucia Dataset|
The trajectory recovered from sequential noisy odometry measurements are shown in red, as more measurements are added (). With every new image, the image is mapped into the appropriate embedding space and subsequently used to query the database for a potential loop-closure. The loop-closures are realized as weak zero rotation and translation relative pose-constraints connecting the query node and the matched node. The recovered trajectories after the pose-graph optimization (in blue) shows consistent long-range, and drift-free trajectories that the vehicle traversed.
V-C Implementation details
Network and Training We take the pre-trained Places205 AlexNet [40, 41], and set all the layers before and including pool5 layer to be fixed, while the rest of the fully-connected layers (fc6, fc7) are allowed to be fine-tuned. The resulting network is used as a base network to construct the Siamese Network with shared weights (See Section IV-B). We follow the distance-weighted sampling scheme as proposed by Wu et al. , and sample 10 times more negative examples as positive examples. The class weights are scaled appropriately to avoid any class imbalance during training. In all our experiments, we set the sampling threshold to 0.9, that ensures that identical places have considerable overlap in their viewing frustums. We train the model for 3000 epochs, with each epoch roughly taking 10s on an NVIDIA Titan X GPU. For most datasets including KITTI and St. Lucia Dataset, we train on 3-5 data sessions collected from the vehicle, and test on a completely new session.
Pose-Graph Construction and Optimization We use GTSAM111http://collab.cc.gatech.edu/borg/gtsam to construct the pose-graph and establish loop-closure constraints for pose-graph optimization. For validating the loop-closure recognition module, the odometry constraints are recovered from the ground truth, with noise injected to simulate dead-reckoned drift in the odometry estimate. They are incorporated as a relative-pose constraint parametrized in with rad rotational noise and m translation noise. We incorporate the loop-closure constraints as zero translation and rotation relative-pose constraint with a weak translation and rotation covariance of m and rad respectively. The constraints are incrementally added and solved using iSAM2  as the measurements are recovered.
In this work, we develop a self-supervised approach to place recognition in robots. By leveraging the synchronization between sensors, we propose a method to transfer and learn a metric for image-image similarity in an embedded space by sampling corresponding information from a GPS-aided navigation solution. Through experiments, we show that the newly learned embedding can be particularly powerful for the task of visual place-recognition as the embedded distances are well-calibrated for efficient indexing and accurate retrieval. We believe that such techniques can be especially powerful as the robot can quickly fine-tune their pre-trained models to their operating environments, by simply collecting more relevant experiences.
- Arandjelovic et al.  R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, and J. Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. In , pages 5297–5307, 2016.
- Bromley et al.  J. Bromley, I. Guyon, Y. LeCun, E. Säckinger, and R. Shah. Signature verification using a” siamese” time delay neural network. In Advances in Neural Information Processing Systems, pages 737–744, 1994.
- Chen et al.  Z. Chen, S. Lowry, A. Jacobson, Z. Ge, and M. Milford. Distance metric learning for feature-agnostic place recognition. In Intelligent Robots and Systems (IROS), 2015 IEEE/RSJ International Conference on, pages 2556–2563. IEEE, 2015.
- Chen et al.  Z. Chen, A. Jacobson, N. Sunderhauf, B. Upcroft, L. Liu, C. Shen, I. Reid, and M. Milford. Deep learning features at scale for visual place recognition. arXiv preprint arXiv:1701.05105, 2017.
- Chopra et al.  S. Chopra, R. Hadsell, and Y. LeCun. Learning a similarity metric discriminatively, with application to face verification. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05), volume 1, pages 539–546. IEEE, 2005.
- Churchill and Newman  W. Churchill and P. Newman. Practice makes perfect? managing and leveraging visual experiences for lifelong navigation. In Robotics and Automation (ICRA), 2012 IEEE International Conference on, pages 4525–4532. IEEE, 2012.
- Cummins and Newman  M. Cummins and P. Newman. Appearance-only SLAM at large scale with FAB-MAP 2.0. The International Journal of Robotics Research, 30(9):1100–1123, 2011.
- Furgale and Barfoot  P. Furgale and T. D. Barfoot. Visual teach and repeat for long-range rover autonomy. Journal of Field Robotics, 27(5):534–560, 2010.
- Galvez-Lopez and Tardos  D. Galvez-Lopez and J. D. Tardos. Bags of binary words for fast place recognition in image sequences. IEEE Transactions on Robotics, 28(5), October 2012. ISSN 1552-3098. doi: 10.1109/TRO.2012.2197158.
Ganin and Lempitsky 
Y. Ganin and V. Lempitsky.
Unsupervised domain adaptation by backpropagation.In
International Conference on Machine Learning, pages 1180–1189, 2015.
- Geiger et al.  A. Geiger, P. Lenz, and R. Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2012.
- Glorot et al.  X. Glorot, A. Bordes, and Y. Bengio. Domain adaptation for large-scale sentiment classification: A deep learning approach. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 513–520, 2011.
- Glover et al.  A. J. Glover, W. P. Maddern, M. J. Milford, and G. F. Wyeth. FAB-MAP + RatSLAM: Appearance-based SLAM for multiple times of day. In Robotics and Automation (ICRA), 2010 IEEE International Conference on, pages 3507–3512. IEEE, 2010.
- Gopalan et al.  R. Gopalan, R. Li, and R. Chellappa. Domain adaptation for object recognition: An unsupervised approach. In Computer Vision (ICCV), 2011 IEEE International Conference on, pages 999–1006. IEEE, 2011.
- Jégou et al.  H. Jégou, M. Douze, C. Schmid, and P. Pérez. Aggregating local descriptors into a compact image representation. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR). IEEE, 2010.
- Jegou et al.  H. Jegou, F. Perronnin, M. Douze, J. Sánchez, P. Perez, and C. Schmid. Aggregating local image descriptors into compact codes. IEEE transactions on pattern analysis and machine intelligence, 34(9):1704–1716, 2012.
- Kaess et al.  M. Kaess, H. Johannsson, R. Roberts, V. Ila, J. J. Leonard, and F. Dellaert. isam2: Incremental smoothing and mapping using the Bayes tree. The International Journal of Robotics Research, 31(2):216–235, 2012.
- Košecká et al.  J. Košecká, F. Li, and X. Yang. Global localization and relative positioning based on scale-invariant keypoints. Robotics and Autonomous Systems, 52(1):27–38, 2005.
- Krizhevsky et al.  A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097–1105, 2012.
- Kuipers and Beeson  B. Kuipers and P. Beeson. Bootstrap learning for place recognition. In AAAI/IAAI, pages 174–180, 2002.
- Kulis et al.  B. Kulis et al. Metric learning: A survey. Foundations and Trends® in Machine Learning, 5(4):287–364, 2013.
- Latif et al.  Y. Latif, C. Cadena, and J. Neira. Robust loop closing over time. Robotics: Science and Systems VIII, page 233, 2013.
- Lowry et al.  S. Lowry, N. Sünderhauf, P. Newman, J. J. Leonard, D. Cox, P. Corke, and M. J. Milford. Visual place recognition: A survey. IEEE Transactions on Robotics, 32(1):1–19, 2016.
- Lynen et al.  S. Lynen, M. Bosse, P. Furgale, and R. Siegwart. Placeless place-recognition. In 3D Vision (3DV), 2014 2nd International Conference on, volume 1, pages 303–310. IEEE, 2014.
- Maaten and Hinton  L. v. d. Maaten and G. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(Nov):2579–2605, 2008.
- Maddern et al.  W. Maddern, M. Milford, and G. Wyeth. CAT-SLAM: probabilistic localisation and mapping using a continuous appearance-based trajectory. The International Journal of Robotics Research, 31(4):429–451, 2012.
- Mei et al.  C. Mei, G. Sibley, and P. Newman. Closing loops without places. In Intelligent Robots and Systems (IROS), 2010 IEEE/RSJ International Conference on, pages 3738–3744. IEEE, 2010.
- Milford  M. Milford. Vision-based place recognition: how low can you go? The International Journal of Robotics Research, 32(7):766–789, 2013.
- Milford and Wyeth  M. J. Milford and G. F. Wyeth. SeqSLAM: Visual route-based navigation for sunny summer days and stormy winter nights. In Robotics and Automation (ICRA), 2012 IEEE International Conference on, pages 1643–1649. IEEE, 2012.
- Oliva and Torralba  A. Oliva and A. Torralba. Building the gist of a scene: The role of global image features in recognition. Progress in brain research, 155:23–36, 2006.
- Philbin et al.  J. Philbin, O. Chum, M. Isard, J. Sivic, and A. Zisserman. Object retrieval with large vocabularies and fast spatial matching. In Computer Vision and Pattern Recognition, 2007. CVPR’07. IEEE Conference on, pages 1–8. IEEE, 2007.
- Russakovsky et al.  O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 2015.
- Simonyan and Zisserman  K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
- Sünderhauf and Protzel  N. Sünderhauf and P. Protzel. Brief-gist-closing the loop by simple means. In Intelligent Robots and Systems (IROS), 2011 IEEE/RSJ International Conference on, pages 1234–1241. IEEE, 2011.
- Sünderhauf et al.  N. Sünderhauf, P. Neubert, and P. Protzel. Are we there yet? challenging SeqSLAM on a 3000 km journey across all four seasons. In Proc. of Workshop on Long-Term Autonomy, IEEE International Conference on Robotics and Automation (ICRA), page 2013, 2013.
- Sunderhauf et al.  N. Sunderhauf, S. Shirazi, A. Jacobson, F. Dayoub, E. Pepperell, B. Upcroft, and M. Milford. Place recognition with convnet landmarks: Viewpoint-robust, condition-robust, training-free. Proceedings of Robotics: Science and Systems XII, 2015.
- Wu et al.  C.-Y. Wu, R. Manmatha, A. J. Smola, and P. Krähenbühl. Sampling Matters in Deep Embedding Learning, 2017.
- Xing et al.  E. P. Xing, A. Y. Ng, M. I. Jordan, and S. Russell. Distance metric learning with application to clustering with side-information. Advances in neural information processing systems, 15:505–512, 2003.
- Zhou et al. [2014a] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Object detectors emerge in deep scene CNNs. arXiv preprint arXiv:1412.6856, 2014a.
- Zhou et al. [2014b] B. Zhou, A. Lapedriza, J. Xiao, A. Torralba, and A. Oliva. In Advances in neural information processing systems, pages 487–495, 2014b.
- Zhou et al.  B. Zhou, A. Khosla, A. Lapedriza, A. Torralba, and A. Oliva. Places: An image database for deep scene understanding. arXiv preprint arXiv:1610.02055, 2016.