The purpose of this document is to provide additional results to complement the main paper. It shows additional renders of a generic bunny model in different scenes, an overview of our new indoor HDR dataset, and all the renders in the user study.

All the provided examples (images and panoramas) are taken from our test set, which the neural network has never seen during training.

Table of Contents

1. Light mask predictions and bunny renders on SUN360 (extends fig. 11)
2. HDR dataset mosaic montage (extends sec. 7.1)
3. HDR prediction on indoor HDR dataset (extends fig. 15)
4. Comparative renders on the HDR dataset used in the user study (extends fig. 17)
5. Illumination scale factor tuning (extends sec. 8.2)

1. Light mask predictions and renders on SUN360

In this section, we present additional predictions from our LDR network on images from the SUN360 dataset. From left to right: the input image, a bunny rendered using an IBL recovered from the network predictions (see sec. 8.1), and the network predictions overlaid on the ground truth warped panorama. Note that this panorama was never seen by the network, nor is it used in the rendering process. These have been randomly sampled from the SUN360 test set, however we took care to include input images that contain appropriate insertion point (that is, avoiding close-ups on objects or completely dark crops).

2. HDR dataset mosaic

In this section, we show an excerpt of our new indoor HDR panorama dataset. The panoramas presented here are LDR images tonemapped from the original HDR files for visualization. The mosaic has been created by randomly sampling 250 panoramas from the dataset. Click for higher resolution. This dataset is freely available for non-commercial use and can be downloaded at indoor.hdrdb.com

3. HDR prediction on indoor HDR dataset

In this section, we present additional results on our HDR dataset (see Fig. 14 for more details), as predicted by our network finetuned on HDR data. The predicted intensity is displayed in log-scale to ease the analysis. Light intensities are color-coded from yellow (high intensity) to blue (low intensity).

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

4. Comparative renders on HDR dataset

In this section, we present additional rendering results on our HDR dataset. In each of these scenes, one or more objects have been inserted. In addition to the input image, we present, in order, the renders produced with the estimation of: the ground truth HDR panorama, our HDR network, our LDR network, [Khan et al. 2006], and [Karsch et al. 2014]. Note that Karsch method failed to produce any result for two of our inputs. Click on the images for higher resolution.

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

5. Global intensity tuning

As shown in Sec. 8.2, network predictions can be further improved by letting an artist set a single intensity scaling parameter. In this section, we show an example of such parameter on the render results. Click on the images for higher resolution.