Transferencia de sentido

This section provides the details of the methodology followed in this work. This includes the choice of HDR to LDR mapping functions, sequence selection, the preparation of mate- rials required for the two experiments and the design and methodology followed to conduct the two subjective experiments. A visual description of the overall work flow is Figure 5.1.

5.2.1 HDR to LDR mapping techniques

Unlike the previous works on tone-mapping evaluation as mentioned earlier in Chapter 4, this work is categorically not a tone-mapping evaluation. Therefore, in this work, three HDR to LDR mapping techniques were chosen such that each represents a different class of HDR to LDR mapping technique and they are as follows:

• A temporally coherent TMO which can also be classified as an SRO (see Section

2.5.6).

• An image appearance model specifically designed for HDR image rendering.

• An alternative technique to extract the optimal exposure from an HDR frame.

The temporally coherent TMO chosen for this work is the Display Adaptive TMO

(mantiuk) proposed by Mantiuk et al. [MDK08] and the details of this TMO has been de-

scribed earlier in Section 2.5.6. The primary reason for choosing this HDR to LDR mapping technique is because it endeavours to reproduce thereferenceHDR sequence with minimal visible distortion and also accounts for temporal coherence (for HDR video sequences), am- bient lighting and target display. In our case the target display was set tolcd-brightin order to exploit the capabilities of the SIM2 HDR display. Also, this TMO in particular performs very well in comparison tests amongst other operators [MBDC14]. A brief overview of this evaluation has been described earlier in Section 4.4.1.

The image appearance model chosen for this work is the iCAM06 HDR image tone compression algorithm proposed by Kuang et al. [KJF07] which is based on the original iCAM framework [MFH∗_{02]. The details of this TMO has been described earlier in Section}

2.5.7. The primary reason for choosing this tone compression algorithm is that it provides an HVS based alternative technique to the multitude of available TMOs and yet at the same time predicts and preserves the colourfulness of the original scene. Moreover, unlike the previous iCAM models, this improved model was designed specifically for HDR image rendering.

As opposed to a transfer function based TMO, Debattista et al. [DBRS∗_{15] pro-}

posed an alternative exposure extraction technique [HW10] which extracts the optimal ex- posure from an HDR frame to fit the maximum possible information from the original HDR data within the allowable bit-depth of 8-bits/pixel/channel. Although this exposure extrac- tion technique has been proposed as a part of an HDR video compression algorithm to create

the base LDR stream (the details of which are described in Section 3.3.10), this technique can also be applied in isolation to map HDR image/video content to an LDR image/video frame. The primary motivation behind selecting this mapping technique is that it provides an alternative technique to a myriad of TMOs (choice of which is very application depen- dent and subjective) to extract the HDR luminance range into a single optimally calculated exposure and maps the exposure into an 8 bit LDR range analogous to an optimally metered 8 bit/pixel/channel image from a camera under varying lighting conditions.

5.2.2 Sequence selection

This section introduces the reader to the HDR video sequences used in this work. Out of a total of 39 HDR video sequences considered, six sequences were shortlisted based on the overall dynamic range of the sequences as well as the source (capture/generation technique) and context of the sequences. A few sequences represented the same scene (same location, same/similar event - different scene cuts) with similar dynamic ranges. In those cases, only one representative sequence was chosen. In other cases, a few sequences were chosen since their overall dynamic range was lower than others (essentially medium dynamic range

approx14₋16-stops) Moreover, it was ensured that the short-listed sequences represent

different capture techniques such as the Spheron VR, Arri Alexa, artificially rendered etc. The shortlisted HDR video sequences (HDRVs), comprising of 150 frames each were so chosen such that the they also represent a wide variety of production techniques. All HDRVs had a resolution of 1920_×1080 and were graded (in absolute luminance terms) such that the pixel values are in the range of 10−4_{to 4000 cd/m}2_.

Figure 5.2 and Table 5.1 provides a brief description of each scene along with a tone mapped frame, overall dynamic range and production technique.

5.2.3 Preparation of materials

Following the selection of three HDR to LDR mapping techniques and six HDRVs, three corresponding LDRVs were created for each of the six HDRVs. The output HDRVs and LDRVs (6 HDRVs + 18 LDRVs = 24 in total) produced were in.hdrformat and in linear RGB colour space. This was necessary since both the HDRVs and LDRVs were subse- quently converted to a SIM2 [SIMa] HDR display suitable mode.

Since, the design of the ranking- and rating based experiments required the use of a single HDR display it was necessary to verify the luminance rating of the displayed sequences. The luminance rating of both the HDRV and LDRV frames were verified using the SpectroDuo PR-680 photo-spectrometer [Pho] and it was ensured that the maximum luminance rating of the HDRVs were within 4000 cd/m2 _{(catered to be within the range}

of the SIM2 display) while the luminance rating of the LDRVs were within 350 cd/m2

(a) Fireplace (b) Welding (c) CGRoom

(d) Jaguar (e) Seine (f) Tears of Steel

Figure 5.2: Short-listed six HDR video sequences

Name Min(Y) Max(Y) DR (stops)

Production technique

Description

Fireplace 10−5 _{4096 25.01 ARRI Alexa An outdoor winter-night scene with a}

bright bonfire in the foreground. Scene post processed.

Welding 0.003 5904 19.85 Spheron VR An indoor scene of a gas welding ma-

chine producing intermittent sparks of very high luminance.

CGRoom 0.001 5008 20.82 Rendered An artificially rendered scene of the

dark basement with an overhead lamp swinging as barrels fall from an over- head shelf.

Jaguar 0.0001 4344 25.30 Canon EOS

1Ds Mark III

An side profile indoor shot of a Jaguar E-Type. Bright lights are placed in the room for artificially expanding the scene dynamic range.

Seine 0.005 8864 20.29 ARRI Alexa Night outdoor scene of the river Seine

in Paris with a brightly lit ferry pro- ducing the high luminance region of the scene. Scene post-processed.

Tears of Steel 0.017 4088 17.62 N.A. A clip extracted from the short film

produced as a part of the Open Movie project by Blender Foundation. Table 5.1: Overview of the scenes used for the rating based psychophysical experiment. Here Min(Y) and Max(Y) refers to the average minimum and maximum luminance of the sequence.

Subsequently, both HDRVs and LDRVs were converted to an custom-built interme- diate video file format (.big) suitable for displaying the HDR video frames at 30 fps on the

SIM2 HDR display.

5.2.4 Hardware and Software resources

Software resources used for both the ranking and the rating based experiment included the 24 video sequences. Hardware resources included a 47" SIM2 HDR display with a 1920_×1080 native resolution, a peak luminance of 4000 cd/m2 _{and a contrast ratio of}

>106_{: 1 [SIMa]. The LDR display used in the experiments was an Alienware 23" IPS}

display, also with a 1920_×1080 resolution, a peak luminance of 350 cd/m2 _{and a maxi-}

mum contrast ratio of 8_×105_{: 1. Further the SpectroDuo photo-spectrometer was used for}

luminance rating verification of both HDRVs and LDRVs.