Real single-material objects from Drexel dataset

As explained in the paper, methods like [20,21,22,23] are not applicable in our case due to their different input and output. Nevertheless, to provide an indicative visual comparison we compare the results of [21] and [20] with HDR inputs and the Singlets variant of our method with LDR inputs on the Drexel dataset. This dataset contains 6 objects photographed in HDR (i.e. 3 exposures were taken) under 5 natural illumination environments with calibrated ground-truth geometry and HDR illumination. Note that, the aforementioned comparison, i.e. [21,20] with HDR inputs vs Singlets with LDR inputs, is the only one possible (as the objects contain a single material we can not run our Triplets variant and all the other combinations presented in the paper) and that the LDR inputs in our case make the problem significantly harder than the HDR ones required as input in [21,20]. The images for [21,20] are directly taken from the corresponding paper; we tried to run the publicly available code of [21] with a grid parameter search for the priors used in the paper (DS-BRDF, gradient and entropy prior), but since we could not get close to the presented results in [21] we limit ourselves to the visual comparison. To this point, we remind that in this paper we are interested in estimating an environment map that does not only have HDR content, but is closer to the truthful illumination. As such, rendering with a new material or under new illumination - which serves well for radiometrically evaluating the recovered illumination and which we do in the video in the Home tab above - would be insufficient in this case as it would hide the structural details of the recovered environment that we are interested for in this work. From the comparison below (similar format as Figure 7 in the paper) it can be seen that we recover more small scale details and overall scene structure over the compared approaches.

Input [21], HDR input
(directly from paper)
[20], HDR input
(directly from paper)
Ours, LDR input
Ground truth