Synthetic DFF Dataset -- CARLA 0.10

Conventional algorithms for depth estimation gener-ally use uncompressed images for training, testing, and validation.Additionally, the increased prominence of synthetic datasetshas, in part, driven significant increases in their size, makingcompression increasingly valuable as a means to reduce storagerequirements. In this work, we apply the DCT-based JPEGand the Wavelet-based JPEG-2000 compression algorithms, witheleven levels of varying compression, to synthetic datasets gener-ated by the CARLA Driving Simulator. Using relatively simpledepth estimation algorithms, which rely on DFF (Depth fromFocus) and DFR (Depth from Reference), this study quan-titatively evaluates the effects of lossy compression on theirperformance. We find that these algorithms are insensitive tolow levels of lossy compression in comparison to the lossless-compressed baselines. Additionally, the JPEG-2000 compressionalgorithm outperforms the more widely used JPEG algorithm forequal compression ratios, indicating that datasets for use in DFFand DFR applications can be compressed using the JPEG-2000algorithm and achieve order-of-magnitude reductions in datasetsize while maintaining comparable accuracy.