From Dusk Till Dawn: Modeling in the Dark

Radenović, Filip; Schönberger, Johannes L.; Ji, Dinghuang; Frahm, Jan‐Michael; Chum, Ondřej; Matas, Jiřı́

doi:10.1109/cvpr.2016.592

Cited by 35 publications

(25 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In this section we briefly summarize the tightly-coupled Bag-of-Words (BoW) image-retrieval and Structure-from-Motion (SfM) 3D reconstruction system [17], [56] that is employed to automatically select our training data. Then, we describe how we use the 3D information to select harder matching pairs and hard non-matching pairs with larger variability.…”

Section: Training Datasetmentioning

confidence: 99%

“…The clustering procedure [18] gives around 20k images to serve as query seeds. The extensive retrieval-SfM reconstruction [56] of the whole dataset results in 1, 474 reconstructed 3D models. Removing overlapping models leaves us with 713 3D models containing more than 163k unique images from the initial dataset.…”

Section: Training Setup and Implementation Detailsmentioning

confidence: 99%

See 1 more Smart Citation

Fine-Tuning CNN Image Retrieval with No Human Annotation

Radenović

Tolias

Chum

2019

IEEE Trans. Pattern Anal. Mach. Intell.

Self Cite

955

1,090

View full text Add to dashboard Cite

Image descriptors based on activations of Convolutional Neural Networks (CNNs) have become dominant in image retrieval due to their discriminative power, compactness of representation, and search efficiency. Training of CNNs, either from scratch or fine-tuning, requires a large amount of annotated data, where a high quality of annotation is often crucial. In this work, we propose to fine-tune CNNs for image retrieval on a large collection of unordered images in a fully automated manner. Reconstructed 3D models obtained by the state-of-the-art retrieval and structure-from-motion methods guide the selection of the training data. We show that both hard-positive and hard-negative examples, selected by exploiting the geometry and the camera positions available from the 3D models, enhance the performance of particular-object retrieval. CNN descriptor whitening discriminatively learned from the same training data outperforms commonly used PCA whitening. We propose a novel trainable Generalized-Mean (GeM) pooling layer that generalizes max and average pooling and show that it boosts retrieval performance. Applying the proposed method to the VGG network achieves state-of-the-art performance on the standard benchmarks: Oxford Buildings, Paris, and Holidays datasets.

show abstract

Section: Training Datasetmentioning

confidence: 99%

Section: Training Setup and Implementation Detailsmentioning

confidence: 99%

Fine-Tuning CNN Image Retrieval with No Human Annotation

Radenović

Tolias

Chum

2019

IEEE Trans. Pattern Anal. Mach. Intell.

Self Cite

955

1,090

View full text Add to dashboard Cite

show abstract

“…An SfM model containing both the database and query images is built and the resulting poses of the query images are used as ground truth [37,59,66]. Yet, this approach again relies on local feature matches and can only succeed if the query and database images are sufficiently similar [55]. The benchmark datasets constructed this way thus tend to only include images that are relatively easy to localize in the first place.…”

Section: Introductionmentioning

confidence: 99%

Benchmarking 6DOF Outdoor Visual Localization in Changing Conditions

Sattler

Maddern

Toft

et al. 2018

2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition

550

729

View full text Add to dashboard Cite

Visual localization enables autonomous vehicles to navigate in their surroundings and augmented reality applications to link virtual to real worlds. Practical visual localization approaches need to be robust to a wide variety of viewing condition, including day-night changes, as well as weather and seasonal variations, while providing highly accurate 6 degree-of-freedom (6DOF) camera pose estimates. In this paper, we introduce the first benchmark datasets specifically designed for analyzing the impact of such factors on visual localization. Using carefully created ground truth poses for query images taken under a wide variety of conditions, we evaluate the impact of various factors on 6DOF camera pose estimation accuracy through extensive experiments with state-of-the-art localization approaches. Based on our results, we draw conclusions about the difficulty of different conditions, showing that long-term localization is far from solved, and propose promising avenues for future work, including sequence-based localization approaches and the need for better local features. Our benchmark is available at visuallocalization.net.

show abstract

“…There are 91,642 training images in the dataset, and 98 cluster images that are the same or almost the same as the test dataset. Through the minimum hash and spatial verification methods mentioned in the clustering process, about 20,000 images are selected as query images, 18,1697 pairs of positive images and 551 training clusters, including more than 163,000 clusters [50] from the original dataset. The dataset contains all images from the Oxford 5k [51] and Paris 6k [52] datasets.…”

Section: Training Datasetsmentioning

confidence: 99%

Relative Distribution Entropy Loss Function in CNN Image Retrieval

Shi

Zhang

et al. 2020

Entropy

View full text Add to dashboard Cite

Convolutional neural networks (CNN) is the most mainstream solution in the field of image retrieval. Deep metric learning is introduced into the field of image retrieval, focusing on the construction of pair-based loss function. However, most pair-based loss functions of metric learning merely take common vector similarity (such as Euclidean distance) of the final image descriptors into consideration, while neglecting other distribution characters of these descriptors. In this work, we propose relative distribution entropy (RDE) to describe the internal distribution attributes of image descriptors. We combine relative distribution entropy with the Euclidean distance to obtain the relative distribution entropy weighted distance (RDE-distance). Moreover, the RDE-distance is fused with the contrastive loss and triplet loss to build the relative distributed entropy loss functions. The experimental results demonstrate that our method attains the state-of-the-art performance on most image retrieval benchmarks.

show abstract

From Dusk Till Dawn: Modeling in the Dark

Cited by 35 publications

References 15 publications

Fine-Tuning CNN Image Retrieval with No Human Annotation

Fine-Tuning CNN Image Retrieval with No Human Annotation

Benchmarking 6DOF Outdoor Visual Localization in Changing Conditions

Relative Distribution Entropy Loss Function in CNN Image Retrieval

Contact Info

Product

Resources

About