Comparison of light field compression methods

An Untrained Neural Network Prior for Light Field Compression

“…4) Quantization: In Table V, we compare different quantization schemes in terms of light field reconstruction performance. Four networks with parameters (c a , c s ) = (15,30), (20,40), (25,50), (30,60) are tested. We take pretrained light field generative models, which corresponds to the line "Without quantization" in Table V, and then apply different quantization schemes.…”

“…Finding effective but compact representations of light fields, that would capture both their spatial and angular redundancy, has therefore become a key challenge for practical use of this technology. This motivated the design of a variety of solutions, ranging from approaches extending HEVC Intra to directly compress the lenslet images [12], [13], to the compression of the set of views as pseudosequences using HEVC [14], [15], or using solutions based on 4D disparity-compensated transforms applied on spatio angular blocks [16], [17], [18], [19], [20]. A comparison of the performance of light field compression schemes using various video coding standards can also be found in [20].…”

“…This motivated the design of a variety of solutions, ranging from approaches extending HEVC Intra to directly compress the lenslet images [12], [13], to the compression of the set of views as pseudosequences using HEVC [14], [15], or using solutions based on 4D disparity-compensated transforms applied on spatio angular blocks [16], [17], [18], [19], [20]. A comparison of the performance of light field compression schemes using various video coding standards can also be found in [20]. Methods using view synthesis have also been proposed in [21] to synthesize all the views from a sparse set of input views, or in [22] where the authors use a linear approximation computed with Matching Pursuit for disparity based view prediction.…”