Retargeting scene-referred images

The images below demonstrate retargeting scane-referred (aka HDR) images to a 200 cd/m² sRGB display. This scenario is equivalent to tone-mapping that simulates the changes in human vision in different luminance conditions.

Two variations of our algorithm are shown:

• Faithful reproduction - accurate simulation of the night vision mechanism.
• Exaggerated visualization - the parameters are exaggerated to show stronger appearance change due to night vision. Such reproduction is perceptually less accurate, but gives stronger "cinematographic" look, which better indicates that the image is seen in night-viewing conditions.

The web page also contains the results for:

• Calibrated Color Appearance Model [Reinhard et al. 2012].
• Multi-scale model of adaptation [Pattanaik et al. 1998].
• Perceptual mal-adaptation model [Irawan et al. 2005].

The images are best seen on a display of about 200 cd/m² peak luminance, closely reproducing the sRGB colour space. The images were generated for a distance of approx. 0.5 m on a 24 inch monitor (angular resolution of 30 pixels per visual degree).

To observe

• For our methods, dark image regions lose details while the details in bright regions retain their contrast. This is because contrast constancy is imperfect only for low luminance levels.
• Large contrast is less affected than small contrast. This is because the contrast loss is due to an additive neural noise. This is best seen in the Nancy cathedral image, where texture on the wall is lost while the sharp edges of the wall maintain their contrast.
• The color-shift due to Purkinje effect is the best seen in Palau bay image.
• Calibrated CAM does not account for night vision. In particular:
• No loss of accuity (detail visibility).
• Hue shift gives greenish instead of bluish cast.
• Images rendered mostly irrespective of absolute luminance levels. "Church-cropped" is a cropped version of the "Church" image. Yet, Calibrated CAM produces very different result for both scenes. Cropped version is visualized as a day-light scenes (full adaptation).
• Multi-scale model of adaptation overpredicts the effect of night vision and does not cope well with high contrast images. In particular:
• Most images are too dark due to excessive dynamic range of the resulting image
• Loss of acuity is too strong.
• Loss of acuity is simulated as blurring instead of loss of detail visibility (refer to "Chuch-cropped" image).
• Strong 'halo' effect due to non-linear multi-scale processing.

Interactive features

Click and hold on a retargetted image to see the best exposure.

Click on the thumbnail icon in the left-bottom corner of an image to see a tone-curve. Click on a tone-curve to hide it.

Zoom out in the browser window (usually CMD + "-") to see smaller images.

The comparison can be restricted to the following methods (uncheck to hide):

Original image: BridgeStudios2

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: atrium_night

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: clergy_seats

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: clock_building

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: desk

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: display_1000

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: driving_simulator

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: garage

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: lldno_pier

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: nancy_cathedral

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: nancy_cathedral_crop

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: nigh_car

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: palau_bay

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: tuebingen_1

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: tuebingen_4

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model

Original image: wreath_building

Our method, faithful reproduction

Our method, exagerrated visualization

Calibrated CAM

Multi-scale model of adaptation

Perceptual mal-adaptation model