GAMa: Cross-view Video Geo-localization

Video geo-localization is a challenging problem with many applications such as navigation, autonomous driving, and robotics [11, 24, 5]. The problem to estimate geo-location of the source of a ground video is also faced by first respondents now and then. To solve this problem, there are two main approaches; same-view geo-localization [19, 17, 3] and cross-view geo-localization [13, 35, 16]. In same-view geo-localization, the query ground image is matched with a street view image from the reference set or gallery. Research in videos is limited to same-view geo-localization where a frame-by-frame approach is followed. This approach relies on availability of ground view images for all the locations which may not be possible always.

In such scenarios, cross-view geo-localization is more useful where the query image is matched with the corresponding aerial or satellite image. However, cross-view geo-localization is a much difficult problem since there is a large domain shift between ground and aerial view. The limited field-of-view in the ground view makes this even harder and it is sometimes difficult even for humans to identify the correct location of a given image or video.

The existing works in cross-view geo-localization follow an image based approach where a ground image is matched with an aerial image [16, 29, 26]. In such an approach, the contextual information available with the video is lost. We focus on the geo-localization of ground videos instead of images to utilize the context, i.e. how the view in one frame is related/located w.r.t. another frame. To the best of our knowledge there are no existing datasets which are publicly available and can be used for this problem. Thus, we have collected a new dataset, named as GAMa (Ground-video to Aerial-image Matching), which contains ground videos with GPS labels and corresponding aerial images. It consists of $\sim$ 1.9M aerial images and 51K ground videos where each video is around 40 seconds long. An example video with representative frames and corresponding aerial image is shown in Figure 1. We will make this dataset publicly available for further research.

We propose GAMa-Net as a benchmark method to solve this problem at clip level where we match every 0.5 second (short clip) from a long video with the corresponding aerial view. A frame-by-frame approach can also be used for video geo-localization, where a 2D convolutional network is used to get the spatial features frame-by-frame. However, following a frame-by-frame approach ignores the rich contextual information available in a video. Considering this, we propose a 3D-convolution based approach for learning the ground level features from a video. The proposed approach is trained using an image-video contrastive loss which tries to match the ground view with the aerial view. This matching provides a clip-level geo-localization.

Next, we propose a hierarchical approach which helps in improving the clip-level geo-localization performance while providing a video-level geo-localization with the help of clip-level predictions. It takes the set of aerial images corresponding to the clips of a long video, and matches them against a larger geographical area. Therefore, it makes use of the contextual information available with the sequence of clips corresponding to a longer video.

We evaluate the proposed approach on GAMa dataset and demonstrate its effectiveness for clip/video geo-localization. We provide an analysis and propose a set of baselines to benchmark the dataset. We make the following contributions:

• •

  A novel problem formulation i.e. cross-view video geo-localization and a large-scale dataset, GAMa, with ground videos and corresponding aerial images. This is the first video dataset for this problem to the best of our knowledge.

• •

  We propose GAMa-Net, which performs cross-view video geo-localization at clip-level by matching a ground video with aerial images using an image-video contrastive loss.

• •

  We also propose a novel hierarchical approach which provides video-level geo-localization and utilizes aerial images at different scales to improve the clip-level geo-localization.

Traditional features of classical computer vision were earlier used for image matching in geo-localization [20, 33]. As deep learning has proven successful in feature learning most of the recent studies have adopted CNN based approach for learning discriminative features for image matching [25]. The problem of image or video geo-localization is solved using either the same view, which is mostly the ground view, or cross-view. Same-view Geo-localization makes use of the large collections of geo-tagged images available online [15, 34, 19, 17, 3, 2]. The problem is approached with the assumption that there is a reference dataset consisting of geo-tagged ground images and there is an image corresponding to each query image. The problem is then solved as image retrieval of the matching reference to determine the location of the query. There is some research in video geo-localization as well which is solved at frame level which is followed by trajectory smoothing [6, 17, 3]. However, a more complete coverage by overhead reference data such as satellite/aerial imagery has spurred a growing interest in cross-view geo-localization.

Cross-view Geo-localization. Most of the recent work adopt CNN based approaches. Several studies have explored CNN architectures for matching ground-level query images to overhead satellite images [28, 26, 36, 12, 8, 29]. Triplet loss is mostly used optimization function in these studies [8, 16]. Certain studies however report better results with contrastive loss[15]. Field-of-view (FOV) also plays an important role in deciding the recall rate and ground panorama is highly accurate as compared to limited FOV [8, 35]. Similarly, videos also contain more visual information as seen through the trajectory of the camera and can be expected to provide more accurate geo-localization as compared to images or frames with similar FOV.

Current works on cross-view geolocalization follow image based approach since the existing datasets only contain image pairs for ground and aerial view [25, 16, 9, 31, 18]. However, some papers do report testing their model on videos using a frame-by-frame approach [9]. Most popular datasets, Cross-View USA (CVUSA) dataset [30], CVACT [13], and Vo et al. [28] contain ground panorama aligned with corresponding aerial image. Recent publications have shown very high recall rate on these datasets while using panoramas however these values are quite low when using limited FOV and unaligned images, i.e. top@1 recall of upto 14% [22, 31]. VIGOR [35] dataset also contains panorama however being unaligned it is more realistic. All these datasets use ground panorama which is not realistic from video geo-localization, as videos have limited FOV, neither do they have time series data required for such training. It is possible to get unaligned images and limited FOV from these datasets however, there is no existing dataset with ground videos and aerial image pairs to solve cross-view geolocalization in videos and the proposed dataset addresses this gap.

Cross-view is also used for fine geo-localization of UAVs or robots. Camera feed (also frame-by-frame) and a known small region/map, of about a mile is used to find a more exact location in the given map [7]. Sometimes a prior is given to estimate the vehicle pose [10]. However, our focus is on coarse geo-localization where the gallery spans over multiple cities.

A frame-by-frame geolocalization of videos is also possible with the proposed dataset where 2D convolutional networks are used to extract the spatial features. However, while following a frame-by-frame approach the contextual information as available from a video is ignored. Fusion of features obtained from 2D-CNN is also possible however it is more memory intensive as compared to a 3D-CNN network. Considering these limitations and challenges, we propose a videos based cross-view geo-localization.

The proposed GAMa (Ground-video to Aerial-image Matching) dataset comprises of select videos from BDD100k [32] and aerial images from apple maps.

Ground video selection: Most of the videos in BDD100k dataset are 40 sec long and usually have GPS label every second. The selection of videos from dataset was based on the range of latitude and longitude for a given video where we use a range parameter $\mu$. We label the GPS points at $n^{th}$ second as $X_{n}$ ($lat_{n}$, $long_{n}$), where the corresponding latitude is $lat_{n}$ and longitude is $long_{n}$ (Figure 1R, Aerial image). Thus, for the whole video we have GPS points as $X_{1}$($lat_{1}$, $long_{1}$), $X_{2}$($lat_{2}$, $long_{2}$),.., $X_{n}$($lat_{n}$, $long_{n}$). If max latitude = $lat_{k}$ and min latitude = $lat_{l}$, then Latitude range = $lat_{k}$ – $lat_{l}$. Also, if max longitude = $long_{p}$ and min longitude = $long_{q}$, then Longitude range = $long_{p}$ – $long_{q}$. Range, $\mu$ = max(Latitude range, Longitude range) In order to eliminate stationary and very fast videos, we select videos with $\mu$ from 0.001 to 0.004. Figure 1R. shows the distribution of training videos with range, $\mu$. The distribution was similar for training and test sets. This selection left us with 46596 training and 6728 testing videos which were further screened based on the availability of GPS labels.

Aerial images: For the selected videos, aerial images are downloaded as tiles from Apple maps at 19x zoom [1]. The dataset comprises of one large aerial image (1792x1792) corresponding to each video of around 40 sec. and 49 uncentered small aerial images (256x256) for these large aerial regions. Table 1 summarizes the dataset statistics. Since, most of the videos have a GPS label every second we divide the videos into smaller clips of 1 sec. each and for each clip we have an aerial tile.

In Figure 1, we see an example of a large aerial region, along with the video frames at each second.

Aerial image centering: In Figure 2(L)a, we show an example of a large aerial image. We have a centered (CN) and an uncentered (UCN) set of small aerial images, Table 1. UCN aerial images are obtained by dividing the large aerial image into smaller tiles. The GPS label in these UCN smaller aerial images can be anywhere besides the center, Figure 2(L)b. There are three labels in the figure and for each of these GPS points the same tile will be considered as the ground truth. For making a centered (CN) set, as shown in Figure 2(L)c we take a crop centered around the corresponding GPS point however it has lead to overlap among the aerial images since in some videos we have a distribution where GPS points are nearby. In the dataset, we still have a one on one correspondence among aerial images and short ground clips. The overlap among the aerial images however increases the difficulty level for top-1 retrieval thus we evaluate with distance threshold.

The dataset covers multiple cities, and the ground videos are distributed all over the US and also from middle eastern region. However, most of the videos are from four US cities; New York, Berkeley, San Francsico, and Bay area. They show different weather conditions, including sunny, overcast, and rainy, as well as different times of the day including day and night.

There is occlusion in videos and shadows of the skyscrapers in aerial images. Limited FOV in videos and all stated characteristics bring it closer to a realistic scenario however also makes it a difficult dataset for geolocalization. In Table 2, we show a comparison with existing datasets for cross-view geolocalization.

An overview of the proposed approach which works on clip and video level is show in Figure 2R. Ground-video to Aerial-image Matching Network (GAMa-Net) learns features for ground view clips and aerial images; and bring the matching pair closer in the feature space by applying a contrastive loss. This provides a clip-level geolocalization for a long video. In addition, we propose an hierarchical approach, where we do video-level geolocalization and use it to update the gallery of aerial images by selecting top matched large aerial regions (Figure 2R). The reduced gallery helps improve the clip-level geolocalization by screening out some of the visually similar however incorrect aerial images.

Implementation and training details: We implement our GAMa-Net and screening network using PyTorch [14] and train using Adam optimizer with a learning rate (lr) of 8e-5. We use a lr scheduler with lr decay rate of 0.1. The screening network is trained in two step; first with the actual ground truth sequences, and then finetuned with predictions from GAMa-Net. The finetuning step allows the network to adapt to noisy aerial sequences which will be used during inference. We use pre-trained weights from Kinetics-400 for 3D-ResNet18 and ImageNet weights for 2D-ResNet18. The ground videos in the proposed dataset are from different times of the day, for faster training we have used only day videos in our experiments unless stated otherwise.

Evaluation: We use top-k recall for clip-level and video-level matching at different values of k. Given a video query, its closest k reference neighbors in the feature space are retrieved as predictions. Similar to image based geolocalization methods we use recall rates at top-1, top-5, top-10, and top 1%. More details can be found in [28, 8, 21]. We have UCN and CN sets of aerial images corresponding to each clip. In UCN set there is one-to-one correspondence between the clip and aerial image. The GPS point can be anywhere within the aerial image however in the CN set there is an overlap among the aerial images. To keep the evaluation similar to UCN set it is considered a correct match if the predicted GPS is within the range of 0.05 mile of the correct location. We also report top-1@t rate where t is a distance threshold to be used for correct prediction.

Baselines: We propose several baselines for comparison. For ground image based baselines we use the center frame of the clip as an input. Image-CBn, our proposed baseline uses 2D-CNN ResNet18 model to encode the ground video frame with similar contrastive loss formulation as GAMa-Net. We use two different loss functions for video based baselines (Triplet-Bn and CBn); margin triplet loss [27] and contrastive loss [4]. In these baselines, we utilize 2D-CNN ResNet18 for aerial images and 3D-CNN ResNet18 for ground videos.

Videos and contextual information: In videos, we have more information which can be considered as contextual information from geo-localization point of view. It enables the network to locate a given frame or view with respect to the other frames in the video. One frame may contain features to complement another and together they are likely to provide more or complete information required for geo-locating. In Table 3, we have compared video based method with frame based baselines. Better recall rate with videos show that the network is able to utilize the additional information available with videos.

Centered vs Uncentered: In an UNC aerial image the corresponding GPS point can be anywhere in the tile. In cases where the GPS point is near the boundary, the visual information from the video is less likely to match the corresponding tile. As expected, after centering (CBn-CN) we observed an improvement in the recall rate as compared with the uncentered set (i.e. CBn-UCN).

Full Dataset: GAMa dataset is a large dataset which has its pros and cons. With large amount of data networks are better trained however this also increases the training time and memory requirement. Here we discuss results with models trained on full dataset (Table 7) i.e. both day and night videos. There is an improvement of 1-4% in recall at all k. Thus, using the hierarchical approach our best R@Top-1 and R@[email protected] is 19.4% and 45.1%, respectively.

Comparison: Our results are comparable to existing image based methods. One recent study reports R@1=13.95% on CVUSA when images are unaligned and have around 70% field of view (FOV) [31]. Most of the cross-view geolocalization datasets such as CVUSA report high R@k while using ground panorama and aligned aerial images which is unrealistic since a normal camera lens has a FOV of around 72% and it is not possible to get aligned images always. In GAMa dataset, the aerial images and ground videos are unaligned, and ground videos have a limited FOV which makes it a more realistic and difficult dataset.

Challenges: In GAMa dataset among the two sets of small aerial images, i.e. CN and UCN, the UCN set is more realistic however difficult for geolocalization. Unaligned aerial images increase this difficulty however, the orientation information can be extracted from the GPS information and is likely to improve the performance [13]. Also, the ground videos have varying lengths and in some of the cases GPS label is not available every second. Thus, the video length available for geolocalization is less than 40 sec. Additionally, there is occlusion because of the car hood and other objects. Such cases are more likely to appear in fail cases.

Limitations: The proposed hierarchical approach performs better with longer videos (8 sec. or more), however it can be used with shorter clips as well. The aim behind using a hierarchical approach is to filter out confusing samples to improve the retrieval rate. However, this sometimes leads to filtering of the ground truth from the gallery. The model is also likely to fail with indoor videos since it will not be possible to match the features with an aerial image. However, this limitation is common to all cross-view geo-localization methods.

In this work, we focus on the problem of video geolocalization via cross-view matching. We propose a new dataset for this problem which has more than 51K ground videos and 1.9 million satellite images. The dataset spans multiple cities and is a more realistic dataset, unaligned and limited FOV. We believe that this dataset will be useful for future research in cross-view video geo-localization. Our proposed GAMa-Net effectively makes use of the rich contextual information available with video. In addition, we propose a hierarchical approach which also utilize the contextual information to further improve the geo-localization.