Perceptual Loss for Robust Unsupervised Homography Estimation

Koguciuk, Daniel; Zonooz, Bahram

doi:10.1109/cvprw53098.2021.00483

Cited by 22 publications

(29 citation statements)

References 24 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Meanwhile, the alignment part of UDIS [44] shares the same multi‐scale network in LB‐DHN [52], so we choose UDIS for testing. biHomE [45] proposes a new method for calculating the loss function, which improves the accuracy of the network in homography estimation and is applicable to scenes with varying illumination conditions. As the best result of all the experiments in this work, we adopt the combination of a fixed pre‐trained feature extractor named Zhang [35] and biHomE loss.…”

Section: Methodsmentioning

confidence: 99%

“…Col 2: The ground truth alignment results. Col 3–8: The alignment results of SIFT [12]+RANSAC [17], DHN [29], Zhang+biHomE [45], VIFSNet [30], UDIS [44] and ours…”

Section: Methodsmentioning

confidence: 99%

“…The network is trained in both supervised and unsupervised ways and a new large real dataset is proposed for further unsupervised image stitching research work. To improve the network robustness to illumination and/or larger viewpoint changes, Daniel et al [45] propose biHomE loss for unsupervised homography estimation, which considering the distance in feature space. This work adopts the fixed pre-training model proposed in previous work as the feature representation stage, and trains the perceptual head stage through biHomE loss, which obtains competitive results on synthetic datasets.…”

Section: Learning-based Homography Estimation Schemesmentioning

confidence: 99%

“…Moreover, to observe the alignment effect of the different methods more intuitively, we also use the predicted matrix to perspective transform the source image and then obtain the coarse stitching as shown in Figure 5. The alignment results of SIFT [12]+RANSAC [17], DHN [29], Zhang+biHomE [45], VIFSNet [30], UDIS [44] and ours The SIFT+RANSAC method uses an identity matrix in the failure cases where SIFT is unable to extract enough feature points, so the stitching results seem to be a plain fusion of the two original images. Similarly, the absolute value of the predicted offset of CA-DHN is so small that the stitching results are roughly seen as a fusion of the original images so we replace the stitching results with those of Zhang+biHomE since they adopt the same backbone.…”

Section: Stitched Mscocomentioning

confidence: 99%

See 3 more Smart Citations

Multi‐scale homography estimation based on dual feature aggregation transformer

Chen

Sun

et al. 2022

IET Image Processing

View full text Add to dashboard Cite

The accuracy of registration in image stitching task directly affects the performance of subsequent stages. Traditional registration methods rely heavily on the quality of the features when calculating the homography matrix, resulting in alignment failures in low‐texture or low‐overlap scenes due to extracting insufficient features. On the other hand, existing DNN‐based methods for homography estimation are more robust in multiple scenes but previous work usually employs an overly simple convolutional network structure to directly regress the homography, ignoring the redundant information contained in the feature maps so that their prediction accuracy is inferior to the traditional methods in simple scenarios. To overcome the disadvantages of the two methods, a Multi‐scale structure is proposed to extract feature maps at three scales and design two modules to handle the matrix prediction respectively. The DFA‐T module analyzes semantic information on the high‐level features to accomplish coarse‐grained alignment while the Contextual Correlation module on the bottom level to accomplish more accurate alignment. Experiments demonstrate that this method provides more accurate alignment results than the existing state‐of‐the‐art DNN‐based methods and outperforms traditional algorithms with more stable results in some extreme scenarios.

show abstract

Section: Methodsmentioning

confidence: 99%

“…Col 2: The ground truth alignment results. Col 3–8: The alignment results of SIFT [12]+RANSAC [17], DHN [29], Zhang+biHomE [45], VIFSNet [30], UDIS [44] and ours…”

Section: Methodsmentioning

confidence: 99%

Section: Learning-based Homography Estimation Schemesmentioning

confidence: 99%

Section: Stitched Mscocomentioning

confidence: 99%

See 2 more Smart Citations

Multi‐scale homography estimation based on dual feature aggregation transformer

Chen

Sun

et al. 2022

IET Image Processing

View full text Add to dashboard Cite

show abstract

“…Traditional methods estimate homography by extracting and matching descriptors between two images, so their performance highly relies on the quality of hand-crafted features [5][6][7][8][9]. On the contrary, deep learning-based methods take two images as input and directly output the homography matrix in an end-to-end fashion [2,10,11]. Although neural networks can employ data augmentation to improve robustness and produce more consistent results [5], they may not always be as accurate as the feature-based methods.…”

Section: Introductionmentioning

confidence: 99%

Learning to Generate High-Quality Images for Homography Estimation

Lin

Zhao

2022

2022 IEEE International Conference on Image Processing (ICIP)

View full text Add to dashboard Cite

Nowadays, robots have gradually replaced humans to perform tasks in challenging environments, but visually satisfactory images may not always contain high-quality features required by downstream algorithms. In this paper, we reveal that in the homography estimation, which is the essential task of robot vision, algorithms show different sensitivities to aerial images and sometimes perform better in over/under-exposed scenes. To this point, we utilize the gamma correction theory to design an interpretable exposure adjustment function, and train a personalized exposure selection network ESH-Net by contrastive learning. After adjusting the exposure of original input by our method, algorithms can obtain better homography results. Comprehensive experiments are conducted on samples generated from the aerial DOTA dataset. Results show that our method can produce high-quality images for both learning-based and traditional homography estimators, and outperforms other exposure enhancement networks.

show abstract