Use of Generative Adversarial Network for Cross-Domain Change Detection

Use of Generative Adversarial Network for Cross-Domain Change Detection

Yamaguchi Kousuke      Tanaka Kanji      Sugimoto Takuma Our work has been supported in part by JSPS KAKENHI Grant-in-Aid for Scientific Research (C) 26330297, and for Scientific Research (C) 17K00361. Graduate School of Engineering, University of Fukui
3-9-1, bunkyo, fukui, fukui
Email: tnkknj@u-fukui.ac.jp
Abstract

This paper addresses the problem of cross-domain change detection from a novel perspective of image-to-image translation. In general, change detection aims to identify interesting changes between a given query image and a reference image of the same scene taken at a different time. This problem becomes a challenging one when query and reference images involve different domains (e.g., time of the day, weather, and season) due to variations in object appearance and a limited amount of training examples. In this study, we address the above issue by leveraging a generative adversarial network (GAN). Our key concept is to use a limited amount of training data to train a GAN-based image translator that maps a reference image to a virtual image that cannot be discriminated from query domain images. This enables us to treat the cross-domain change detection task as an in-domain image comparison. This allows us to leverage the large body of literature on in-domain generic change detectors. In addition, we also consider the use of visual place recognition as a method for mining more appropriate reference images over the space of virtual images. Experiments validate efficacy of the proposed approach.

I Introduction

This paper addresses the problem of change detection from a novel perspective of domain adaptation. Change detection aims to identify interesting changes between a given query image and a reference image of the same scene taken at different times. This is a fundamental problem in computer vision and robotics with many important applications including visual navigation, novelty detection, surface inspection, and city model maintenance. This problem becomes a challenging one when query and reference images involve different domains (e.g., time of the day, weather, season). This is due to variations in object appearance in different domains, which makes it harder to discriminate changes of interest from nuisances. One of most basic schemes to handle this difficulty is to train a domain-specific change predictor for each possible domain pair. However, this requires a vehicle to collect and learn from a number of training images from both domains, which severely limits its applicability to new domains where large amounts of annotated training data are typically unavailable.


Fig. 1: In our approach, a generative adversarial network (GAN) -based image translator is trained and used to map a reference image (left panel) to a virtual one (middle panel) that cannot be discriminated from the query domain images. This enables us to treat the cross-domain change detection as an in-domain image comparison between the query (right panel) and the virtual reference (middle panel) images.

In this study, we address the above issue by leveraging a generative adversarial network (GAN) [1] (Fig. 1). Our approach is motivated by the recent success of the image-to-image translation technique in the field of computer vision and graphics. The image-to-image translation technique aims to learn the mapping between an input image in one domain to a virtual image that cannot be discriminated from target domain images. The key concept we employ is to use a limited amount of training data to train a GAN-based image translator for reference-to-query domain translation. This enables us to treat the cross-domain change detection task as an in-domain image comparison and allows us to leverage the large body of literature on in-domain generic change detectors. In addition, we also consider the use of visual place recognition as a method for mining more appropriate reference images over the space of virtual images. Experiments validate efficacy of the proposed approach.

Ii Related Work

Change detection has been widely studied in various task scenarios. These scenarios are categorized into 2D-to-2D matching [2], 3D-to-3D matching [3], 2D-to-3D matching [4]. [2] considers a 2D-to-2D matching on overhead imagery using an image registration and scale invariant feature transform (SIFT). [3] addressed a 3D-to-3D matching on point cloud data for novelty detection by patrol robots. [4] considered a 2D-to-3D matching between monocular image and cadastral 3D city models. Our approach belongs to 2D-2D matching, and our setting is far more complicated than the case of overhead imagery [2] as our vehicular applications require the capability to deal with more general 6 degrees-of-freedom (DoF) freely-moving camera motion in a 3D space.

Up until now, the most fundamental scheme reported for addressing this challenge is to directly compare each view image against the corresponding reference image. In [5], a scene alignment method for image differencing is proposed based on ground surface reconstruction, texture projection, image rendering, and registration refinement. In [6], a deep deconvolutional network for pixelwise change detection was trained and used for comparing query and reference image patches. However, a major concern of these direct methods is that they may fail to capture appearance variations caused by a domain shift.

Our approach is inspired by the domain adaptation and transfer learning approaches, ranging from parameter adaptation, feature transformation, and metric learning, to deep learning techniques, which have been applied to a wide variety of visual recognition tasks [7]. In recent years, the application of GANs to domain adaptation tasks has attracted much attention [1]. In our application domain, image-to-image translation has been one of mostly studied issues over the last two years [8]. Image-to-image translation aims to map a raw image in a source domain to a virtual image in such a way that the virtual image cannot be discriminated from raw images in the target domain. Typical training algorithms for the translation GAN take a set of paired source-target images as input. We address this basic setting and use a typical architecture in [8] for the current experimental system.


Fig. 2: Experimental environments and viewpoints. The red circles and blue boxes indicate the viewpoints of the 100 pairs of summer and winter images in the test set.

Iii Problem: Cross-Domain Change Detection

Although our approach is sufficiently general and applicable to various types of environments (indoor and outdoor for instance) and sensor modalities, in our experiments, we focus on the NCLT dataset [9]. The NCLT dataset is a large scale, long-term autonomy dataset for robotics research collected at the University of Michigan’s North Campus by a Segway vehicle robotic platform. The Segway was outfitted with: a Ladybug3 omnidirectional camera, a Velodyne HDL-32E 3D lidar, two Hokuyo planar lidars, an IMU, a single axis fiber optic gyro (FOG), a consumer grade global positioning system (GPS), and a real-time kinematic (RTK) GPS. The data we used in the research includes image and navigation data from the NCLT dataset. The image data is from the front facing camera (camera#5) of the Ladybug3 omnidirectional camera. Fig. 2 shows a bird’s eye view of the experimental environment and the vehicle’s camera viewpoints.

During the vehicle’s trips through the outdoor environments (Fig. 2), it encounters various types of changes, which originate from the movement of individuals, parking of cars, building construction, and opening or closing of doors. There are also nuisances that originate from illumination alterations, viewpoint-dependent changes of objects’ appearances and occlusions, weather variations, and falling leaves and snow. A critical and significant challenge in a substantial majority of change detection tasks is to discriminate changes of interest from nuisances. This renders our change detection task significantly more demanding.

Precision-recall is a standard performance index for image based change detection algorithms [10]. As a performance index, we use mean average precision (mAP), derived from the field of visual object detection [11]. A change detection algorithm outputs a prediction of the likelihood of change (LOC) for each cell on a image grid of LOC values, as we will describe in IV-C. For the evaluation, the LOC value of each cell in the image grid is thresholded into a state of change or no-change, and then the binary change mask is compared against a ground-truth change mask. Different threshold values provide different tradeoff points between precision and recall. The mAP can be viewed as a summary of these different tradeoff points and approximates the value of receiver operating characteristic (ROC). Following the literature [11], we approximate the average precision (AP) based on the average of “interpolated” precision values for 11 different recall values, 0, 10, …, 100 [%]. The mAP is obtained as the mean of the AP values over all the query images. Considering the fact that achieving 100% recall is not necessarily required in typical change detection applications, we also evaluate mAP@%recall for different percentage points 10, 20, 50, and 100.

Iv Approach: Domain-Adaptive Change Detection

Our domain-adaptive change detection (DA-CD) framework consists of three distinct steps: image adaptation, image description and image comparison. Image adaptation aims to map a given reference (or background) image to a virtual image, using the GAN-based image-to-image translation, in such a way that the virtual reference image cannot be discriminated from raw images in the target domain. Image description aims to extract a set of image feature descriptors from either image for image comparison. Image comparison aims to predict a change mask by comparing local features between the query and reference images. The three steps above are detailed in the following subsections.

Iv-a Image Adaptation

In order to train a GAN that translates an image from the source domain to the corresponding virtual images in the target domain, we require a small number of paired source-target images as training data. We pair images based on their viewpoint locations. The task of pairing requires global viewpoint information in the source and target domains. Fortunately, in vehicular applications, it is often straightforward to obtain the small set of training images and the aforementioned global viewpoint information. First, a small number of training sets can be acquired by the vehicle’s short range navigation in the target domain. Second, if the GPS measurement of the viewpoints are available, a target image with nearest neighbor viewpoint is selected as the corresponding image. Even in a GPS-denied environment, pseudo GPS information can be obtained with sufficiently high quality by using a modern visual place recognition (VPR) technique. For example, in [12], we also have developed a robust approach to cross-season VPR.

The GAN-based translation problem is often formulated as a per-pixel regression problem. However, this formulation treats the output space as “unstructured” in the sense that each output pixel is considered conditionally independent from all others given the input image. This issue is addressed by conditional GAN with a structured loss function, which penalizes a joint configuration of the output. In particular, the approach proposed in [8] aims for a general-purpose solution to image-to-image translation problems. The proposed approach consists of a generator based on U-Net which allows low-level information to take a shortcut across the network, and a discriminator based on PatchGAN that combines a typical L1 loss (that enforces correctness at low-frequencies) with an additional patch level loss (that models high-frequencies). We also refer to the results in [8], where the approach was found to be effective for a domain adaptation task of “day-to-night” image translation, which is partly similar to the application of a cross-season image translation task that we present. In both the training and translation tasks, we resize each input image to 256256.

Iv-B Image Description

We follow the literature (e.g., [13]) and use the standard local image feature descriptor, SIFT [14], as the basis for image comparison and change detection. In this study, we are particulary interested in how effective intensity-non-invariant feature descriptors like SIFT descriptors are when combined with our strategy of image adaptation. Intuitively, when the image adaptation is successful, the image comparison between the query image and the adapted background image can be viewed as an in-domain image comparison task rather than a cross-domain image comparison. Therefore, intensity-non-invariant feature descriptors are expected to be effective for such a pseudo in-domain image comparison. Each SIFT feature descriptor is L1 normalized.

We basically use a Harris-Laplace detector as a feature keypoint detector. This detector typically provides a set of well-localized keypoints. However, in low texture image regions, it fails to detect localized keypoints and the spatial density of keypoints tend to be very low. To compensate, we also employ an alternative dense sampling -based detector. The dense sampling -based detector outputs a large number of spatially uniform keypoints even in a low texture region. However, as a downside, such keypoints from dense sampling tend to be not well-localized and have a low invariance. This potentially yields a bias due to an increase in dissimilarity between local features at these dense keypoints, which then increases the false positive change detection. To deal with this issue, we decided to penalize the LOC values of these features by multiplying a pre-defined coefficient ().

By default, the features from the virtual image are merged with those from the raw reference image, and then the merged set is used for the image comparison. In the experiments, we also consider an alternative setting where the virtual image set alone is used (without the merge) for image comparison. The above two strategies are respectively termed “w/ merge” and “w/o merge”.

Iv-C Image Comparison

The image comparison step adopts the nearest neighbor (NN) search of local features. In offline mode, every feature in the reference image is indexed by an NN data structure. Once online, the NN search is performed for each query feature over the reference (raw/virtual) features. The search result provides a dissimilarity score of a query feature’s NN reference features with respect to the query feature, which can be interpreted to the LOC score of the query feature. We measure the dissimilarity by calculating the Euclidean distance between query feature and its NN feature . The search region on the image plane is defined as a radius 10-pixel circular region centered at the query keypoint of interest, for 256256 query and reference images.

We also adopt a feature matching -based feature penalization. This strategy has been found to be effective in our previous research in [15]. The basic function of the feature matching approach is to conduct a similarity search for a given query feature over a collection of spatially relevant reference features. If a strong match is found, we can expect the actual possibility of the feature being changed to be very low. Therefore, we penalize the LOC values of these features by multiplying by pre-defined coefficient (). To detect strong matches, we use the SIFT ratio test [16] with a threshold value of 0.4. In preliminary experiments, we also tested different threshold values ranging from 0.4 to 0.8 and found that the difference among these threshold values yield a low significance in change detection performance. The feature matching approach is expected to effectively suppress false positive change detection even when ambiguous anomaly features exist.

In typical change detection applications, we are interested in pixel-level LOC, rather than feature-level LOC [10]. To obtain such pixel-level LOC values, we introduce a coarse image grid of LOC values, where the width and height of the grid is set to and for a image. The LOC value of each grid cell is computed from the LOC values of query features that belong to that grid cell by using max pooling.

In addition, we also explored the use of the VPR technique as a method for selecting appropriate background images, which we call a “background mining” strategy. In general, a virtual image generated by a GAN-based image-to-image translation is often inconsistent with the corresponding query image. This is natural because the query image is unseen and not available at the GAN training stage. To address this issue, here we use the VPR technique as a method for searching for a more relevant background image to a given query image, over the space of all the virtual images of the target domain. In this paper, the VPR algorithm adopts a local feature NN search with a Naive Bayes nearest neighbor (NBNN) distance metric, which was found to be effective in a cross-domain VPR task in our previous research [12]. In experiments, we consider and compare two distinct cases: one for the normal background image output by the GAN-based image-to-image translator, and one for the virtual background image selected by background mining.


b
Fig. 3: mAP@[%]recall performance for different strategies.

a

b

c

d

e

f

g

h

i

j
Fig. 4: Examples of change detection. From left to right, each panel shows the input query image, prediction (red boxes) and ground-truth (blue points) change detection overlaid on the query image, a virtual reference image, and the input reference image. The prediction results (red boxes) correspond to image grid cells with the top 5% LOC values.

V Experiments

We evaluated the change detection strategies using an NCLT dataset [9]. As described in III, the NCLT dataset is a large scale, long-term autonomy dataset for robotics research collected at the University of Michigan’s North Campus by a Segway vehicle robotic platform. During the vehicle’s trip through outdoor environments, it encounters various types of interesting changes and nuisances. A critical and significant challenge in a change detection task is to discriminate changes of interest from nuisances. We consider two different scenarios of domain-shift: winter-to-summer and summer-to-winter. Our winter image set is generated from the datasets “2012/1/22”, “2012/1/15”, and “2013/1/10” in NCLT, while our summer set is generated from “2012/8/4”, by the following procedure. First, images from different seasons are paired using the GPS information available in the NCLT dataset. Then, a random subset of the paired images is used for training the GAN-based image-to-image translator. Two translators corresponding to winter-to-summer and summer-to-winter translation tasks are trained using 100 and 100 paired training images, which are independent from the test sets and randomly sampled from the both domains. Then, every paired image is manually examined and annotated with a bounding box of a change object. We obtain 100 annotated image pairs as the test set, which consist of 64 summer-to-winter pairs and 36 winter-to-summer pairs.

We categorized small changes (e.g., 1010 [pixels]) that typically originate from distant objects, into no-change because it is challenging to detect such small changes via a visual change detection algorithm. As a result, such small objects are likely to cause pseudo false-positive detection by change detection algorithms.

The average time-overhead for image translation was around 1.0 sec and the time-overhead for change detection was 1.0 sec on a single PC (Intel Core i7-7700 3.2 GHz, GeForce GTX 1080).

Fig. 3 shows the mAP@%recall performance for different change detection strategies. As mentioned in IV-B, we consider two settings, whether or not features from the virtual reference images are merged with features from the raw reference image, which are respectively called “w/ merge” and “w/o merge”. In addition, we consider alternative settings and whether or not we use the background mining described in IV-C, which is termed “w/ mining” and “w/o mining”. Figure 3 shows the mAP performance for different combinations of these three strategies. It can be observed that the combination “merge+mining” yields the highest mAP@100%recall performance.

Fig. 4 displays examples of change detection results. It can be seen that in these examples, the virtual reference image preserves the color and structure information and successfully generates a realistic virtual image, while the edge information is often blurred or lost. It is noteworthy that even with such erroneous edge information, the change detection was successful as reported in Fig. 3.

Vi Conclusions

In this paper, we discussed the use of a generative adversarial network (GAN) in cross-domain change detection tasks. We proposed a novel framework of domain adaptive change detection (DA-CD) to capture the domain shift in the images, and combined the local feature search with the GAN-based image translation to capture spatial inconsistencies and appearance inconsistencies between different domains. Improved detection performance was achieved on the challenging cross-season NCLT dataset.

Acknowledgment

Our work has been supported in part by JSPS KAKENHI Grant-in-Aid for Scientific Research (C) 26330297, and for Scientific Research (C) 17K00361.

References

  • [1] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,” in Advances in Neural Information Processing Systems 27, Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, Eds.   Curran Associates, Inc., 2014, pp. 2672–2680.
  • [2] W. Li, X. Li, Y. Wu, and Z. Hu, “A novel framework for urban change detection using vhr satellite images,” in Proceedings of the 18th International Conference on Pattern Recognition-Volume 02.   IEEE Computer Society, 2006, pp. 312–315.
  • [3] P. Drews, P. Núñez, R. Rocha, M. Campos, and J. Dias, “Novelty detection and 3d shape retrieval using superquadrics and multi-scale sampling for autonomous mobile robots,” in Robotics and Automation (ICRA), 2010 IEEE International Conference on.   IEEE, 2010, pp. 3635–3640.
  • [4] A. Taneja, L. Ballan, and M. Pollefeys, “City-scale change detection in cadastral 3d models using images,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2013, pp. 113–120.
  • [5] D. W. van de Wouw, G. Dubbelman, and P. H. de With, “Hierarchical 2.5-d scene alignment for change detection with large viewpoint differences,” IEEE Robotics and Automation Letters, vol. 1, no. 1, pp. 361–368, 2016.
  • [6] P. F. Alcantarilla, S. Stent, G. Ros, R. Arroyo, and R. Gherardi, “Street-view change detection with deconvolutional networks.” in Robotics: Science and Systems, 2016.
  • [7] V. M. Patel, R. Gopalan, R. Li, and R. Chellappa, “Visual domain adaptation: A survey of recent advances,” IEEE signal processing magazine, vol. 32, no. 3, pp. 53–69, 2015.
  • [8] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” arxiv, 2016.
  • [9] N. Carlevaris-Bianco, A. K. Ushani, and R. M. Eustice, “University of michigan north campus long-term vision and lidar dataset,” The International Journal of Robotics Research, vol. 35, no. 9, pp. 1023–1035, 2016.
  • [10] R. J. Radke, S. Andra, O. Al-Kofahi, and B. Roysam, “Image change detection algorithms: a systematic survey,” IEEE transactions on image processing, vol. 14, no. 3, pp. 294–307, 2005.
  • [11] M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman, “The pascal visual object classes (voc) challenge,” International journal of computer vision, vol. 88, no. 2, pp. 303–338, 2010.
  • [12] K. Tanaka, “Cross-season place recognition using nbnn scene descriptor,” in Intelligent Robots and Systems (IROS), 2015 IEEE/RSJ International Conference on.   IEEE, 2015, pp. 729–735.
  • [13] J. Košecka, “Detecting changes in images of street scenes,” in Asian Conference on Computer Vision.   Springer, 2012, pp. 590–601.
  • [14] K. E. A. van de Sande, T. Gevers, and C. G. M. Snoek, “Empowering visual categorization with the gpu,” IEEE Transactions on Multimedia, vol. 13, no. 1, pp. 60–70, 2011.
  • [15] T. Murase, K. Tanaka, and A. Takayama, “Change detection with global viewpoint localization,” in Pattern Recognition (ACPR), 2017 4th IAPR Asian Conference on.   IEEE, 2017, pp. 31–36.
  • [16] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” International journal of computer vision, vol. 60, no. 2, pp. 91–110, 2004.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
""
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
   
Add comment
Cancel
Loading ...
14114
This is a comment super asjknd jkasnjk adsnkj
Upvote
Downvote
""
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters
Submit
Cancel

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test
Test description