Abstract:
Stitching multiple images for achieving the 360 view of any environment is a challenging task. Traditionally, the whole process of image stitching is based on distinctive features that are very helpful for estimating the other parameters of the whole algorithm. As different images require different suitable parameters or weights for achieving the best results and we need to predict those suitable parameters for each case independently. In our proposed model first small neural network based techniques are implemented that are just used for estimating the quality panorama hyper parameters and then we apply the whole stitching algorithm on sample images by using those predicted parameters.
Therefore, due to lack of labeled data we are unable to train any supervised model for those hyper parameter selection that’s why we build an unsupervised technique that makes decisions based on just extracted features quality, confidence and count of inliers etc.
By estimating the good parameters we are able to stitch a quality panorama that doesn't have any ghosting artifacts, blending discontinuities, seamless and alignment errors as well. We evaluate the performance of our proposed model on three datasets and analyze performance in both perspective quality and computational time and conclude that our model outperforms with other state of the art stitching algorithms in both perspectives
