Learning Search Path for Region-level Image Matching

Krishna, Onkar; Irie, Go; Wu, Xiaomeng; Kawanishi, Takahito; Kashino, Kunio

doi:10.1109/icassp.2019.8682714

Cited by 2 publications

(10 citation statements)

References 13 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Quantitative Results. Success rate (higher is better) vs. number of search steps (lower is better) for ours and the baseline method [9]. Ours achieves higher success rate with a fewer number of steps compared to the baseline.…”

Section: Methodsmentioning

confidence: 87%

“…Baselines. We compared our approach with the existing deep learning-based image matching method [9] as well as two handcrafted image matching approaches, MTM [15] and BBS [16]. We used the codes provided by the authors, ensuring that all approaches used the same hardware environment and carefully tuned hyperparameters.…”

Section: Methodsmentioning

confidence: 99%

“…For matching, several methods have been proposed to learn correspondences between images [5][6][7][8]. Unlike these two-step approaches, a more recent study [9] proposed to directly predict the region in the reference image that is likely to match the query. This method uses a recurrent neural network that sequentially explores only highly prospec-tive regions on the reference image, and the network is trained by reinforcement learning in an end-to-end manner.…”

Section: Introductionmentioning

confidence: 99%

“…This method uses a recurrent neural network that sequentially explores only highly prospec-tive regions on the reference image, and the network is trained by reinforcement learning in an end-to-end manner. While this approach can achieve remarkable efficiency by bypassing the time-consuming feature similarity computations, one drawback of the method [9] is that the algorithm itself has no mechanism to know whether the target region is correctly localized or not; even if it succeeds in finding a matching region in a few steps, the search is always continued until it reaches the predetermined maximum number of search steps, thus failing to provide the inherently achievable efficiency.…”

Section: Introductionmentioning

confidence: 99%

“…In this paper, we propose a new image matching network equipped with a self-termination mechanism that can determine whether a region matching the query has been found or not and stop the matching process by itself. Our network overall is designed to sequentially predict and search for regions in the reference image that can match the query as [9]. Our self-terminating mechanism is implemented as a halting module that classifies whether a query and a reference region match or not based on their features and the search history encoded into a short vector.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Deep Reinforcement Image Matching with Self-Termination

Krishna

Irie

et al. 2021

2021 IEEE International Conference on Image Processing (ICIP)

Self Cite

View full text Add to dashboard Cite

Deep reinforcement learning-based image matching sequentially searches only the promising regions in the reference image that match the query, leading to a significantly small number of steps compared to traditional methods. Since existing methods do not have any function to judge whether the target region has been successfully identified or not, they continue to search until the preset maximum number of search steps is reached. In this paper, we propose a deep image matching network that can terminate the matching process by itself. Our network is designed to have a halting module that identifies whether the current reference region matches the query based on the image features and the search history. The entire network is effectively trained end-to-end in a framework of deep reinforcement learning that incorporates a new loss function to evaluate the accuracy of the termination decision. Experimental results demonstrate that our method can achieve highly competitive or better matching accuracy with fewer search steps than the existing methods.

show abstract

Section: Methodsmentioning

confidence: 87%