A Fast, Scalable, and Reliable Deghosting Method for Extreme Exposure Fusion

Prabhakar, K. Ram; Arora, Rajat; Swaminathan, Adhitya; Singh, K. B.; Babu, R. Venkatesh

doi:10.1109/iccphot.2019.8747329

Cited by 46 publications

(56 citation statements)

References 18 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…It is very tedious to collect large labeled datasets for different sequence lengths, as it requires capturing both static and dynamic sequences of same scene. Prabhakar et al [20] addressed scalability in their work by concatenating the mean and max of input features. However, their method still suffers from artifacts in challenging scenarios (shown in Fig.…”

Section: Proposed Methods a Methods Overviewmentioning

confidence: 99%

“…Yan et al [60] use a network with a non-local module to identify matching neighbor features to fill in ill-exposed regions of the reference image. Prabhakar et al [20] aggregate features derived by shared CNN modules to create a scalable architecture that can fuse an arbitrary number of images without retraining. Prabhakar et al [22] propose a method to minimize the consumption of memory and enable the fusion of high-resolution images.…”

Section: Related Workmentioning

confidence: 99%

“…[18] dataset consists of 74 training and 15 validation sequences, while [20] dataset consists of 466 training and 116 validation sequences. In both [18] and [20] datasets, each sequence consists of three varying exposure images with exposure bias {−2, 0, +2} or {−3, 0, +3}. Additionally, we trained on Prabhakar et al [78] dataset for validating scalability performance.…”

Section: A Implementationmentioning

confidence: 99%

“…1). The CNN-based method proposed by Prabhakar et al [20] is capable of fusing a variable number of images, but at a considerable cost in terms of quality due to their internal aggregation strategy of features (see Fig. 7c and 10).…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Self-Gated Memory Recurrent Network for Efficient Scalable HDR Deghosting

Kathirvel¹,

Agrawal²,

Radhakrishnan³

2021

IEEE Trans. Comput. Imaging

View full text Add to dashboard Cite

We propose a novel recurrent network-based HDR deghosting method for fusing arbitrary length dynamic sequences. The proposed method uses convolutional and recurrent architectures to generate visually pleasing, ghosting-free HDR images. We introduce a new recurrent cell architecture, namely Self-Gated Memory (SGM) cell, that outperforms the standard LSTM cell while containing fewer parameters and having faster running times. In the SGM cell, the information flow through a gate is controlled by multiplying the gate's output by a function of itself. Additionally, we use two SGM cells in a bidirectional setting to improve output quality. The proposed approach achieves state-of-the-art performance compared to existing HDR deghosting methods quantitatively across three publicly available datasets while simultaneously achieving scalability to fuse variable length input sequence without necessitating re-training. Through extensive ablations, we demonstrate the importance of individual components in our proposed approach. The code is available at https://val.cds.iisc.ac.in/ HDR/HDRRNN/index.html.

show abstract

Section: Proposed Methods a Methods Overviewmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: A Implementationmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Self-Gated Memory Recurrent Network for Efficient Scalable HDR Deghosting

Kathirvel¹,

Agrawal²,

Radhakrishnan³

2021

IEEE Trans. Comput. Imaging

View full text Add to dashboard Cite

show abstract

“…However, failing this assumption such as in case of deformable-body motions with overlapping regions between exposures, dynamic background in the scene, or insufficient number of input exposures can lead to poor de-ghosting quality. Moving object reconstruction methods [20][21][22][23][24][25][26][27][28][29][30][31][32][33] select one or more reference images for the regions affected by motion. The moving objects are then reconstructed or replaced by the same moving objects extracted from the reference images.…”

Section: Introductionmentioning

confidence: 99%

Untitled

2020

Adv. Electr. Comp. Eng.

View full text Add to dashboard Cite

A High Dynamic Range (HDR) image produced from a sequence of low dynamic range (LDR) images can contain motion artefacts (ghosting) if the scene contains moving objects. Conventional de-ghosting methods first detect moving objects in the scene, and then either remove those moving objects totally or reconstruct them. However, these methods are computationally expensive. This paper proposes a de-ghosting method that does not require explicit detection of moving regions. First, the ratio between camera exposure times of a target image and a reference image, which is called the intensity scaling factor in this paper, is computed. Since the information about camera exposure time is not available always, we propose a novel method to estimate the intensity scaling factor from non-saturated and non-moving pixels. Then, the estimated scaling factor is used as a cue to label every pixel in the target image as either static or moving pixel. Finally, the values of moving pixels are corrected with their expected values which can be estimated from the intensity scaling factor. Experimental results show that the proposed method generates more accurate ghost-free HDR images than the existing state of the art methods.

show abstract