Supplementary Materials Towards Photography Through ...guysatat/fog/materials/... · ing a...
3
Supplementary Materials Towards Photography Through Realistic Fog Guy Satat, Matthew Tancik, Ramesh Raskar MIT Media Lab [email protected] 1. Description of Videos 1. Overview.mp4 This is an overview video provid- ing a high-level description of the paper and results. 2. Es Over Time.mp4 Shows the reconstruction for the ’E’s targets as fog is being added into the chamber. The video starts with an empty chamber, and ends when the fog level in the chamber saturates. The video shows the PointGrey camera measurements, SPAD photon counts, SPAD time gating, and our re- flectance and depth reconstruction. Similarly to Fig. 5 in the main text, we provide SSIM and PSNR for quan- titative comparison. Each frame indicates the mea- sured optical thickness. 3. Mannequin Close Over Time.mp4 Similar to the above for the mannequin measurements, when the mannequin is 35cm away from the camera. 4. Mannequin Far Over Time.mp4 Similar to the above for the mannequin measurements, when the mannequin is 65cm away from the camera. 5. Optical Thickness Over Time.mp4 Shows the PointGrey camera perspective for a scene composed of posts separated by 5cm and starting at 27cm away from the camera. Each frame shows the measured op- tical thickness. 2. Further Experimental Details Supplemental Fig. 1 shows the dimensions of the ‘E’ shape target. Our reconstructions in the main text demon- strate recovery with 1cm spatial resolution and 5cm axial resolution for targets at a range of 37-57cm. 3. Extension for Figure 3 from the Main Text Supplemental Figs. 2+3 provide results on top of Fig. 3 in the main text. Supplemental Fig. 2 shows results that correspond to the left column of Fig. 3 in the main text 2cm 1cm 1cm 1cm 1cm 1cm 1cm Figure 1. Sketch of the E target. The actual target is composed of four identical E shapes separated by 5cm from one another. (estimated KDE, Gamma distribution, signal distribution, and target distribution). Supplemental Fig. 3 shows results that correspond to the right column of Fig. 3 in the main text (histogram and fitted model from Eq. 5). Both figures show results for targets at five different depths (42, 47, 52, 57, 62cm), and a background pixel. Both figures show results for optical thicknesses (OT) of: 0, 1.0, 1.6, 1.9, 2.3, 2.8. All plots are normalized to [0,1], and the title of each subplot indicates the total photons measured. We note that when the object is closer, it can be recov- ered with more challenging OT. For example the closest ob- ject at 42cm is always correctly recovered, the next target at 47cm is correctly recovered up to OT=2.3, and then fails for OT=2.8, and the furthest target (62cm) fails as early as OT=1.6. We also note that for the background, as soon as a little fog is added, the number of counts rises very rapidly, and then stays roughly the same for denser fog (the response shape does change as more fog is added). The opposite pro- cess occurs for the targets – as more fog is added, we note less counts. Lastly, supplemental Fig. 3 shows that the suggested model captures the measurements well at diverse levels of fog (OT) and target depths.
Transcript of Supplementary Materials Towards Photography Through ...guysatat/fog/materials/... · ing a...
Supplementary MaterialsTowards Photography Through Realistic Fog
Guy Satat, Matthew Tancik, Ramesh RaskarMIT Media [email protected]
1. Description of Videos1. Overview.mp4 This is an overview video provid-
ing a high-level description of the paper and results.
2. Es Over Time.mp4 Shows the reconstructionfor the ’E’s targets as fog is being added into thechamber. The video starts with an empty chamber,and ends when the fog level in the chamber saturates.The video shows the PointGrey camera measurements,SPAD photon counts, SPAD time gating, and our re-flectance and depth reconstruction. Similarly to Fig. 5in the main text, we provide SSIM and PSNR for quan-titative comparison. Each frame indicates the mea-sured optical thickness.
3. Mannequin Close Over Time.mp4 Similar tothe above for the mannequin measurements, when themannequin is 35cm away from the camera.
4. Mannequin Far Over Time.mp4 Similar to theabove for the mannequin measurements, when themannequin is 65cm away from the camera.
5. Optical Thickness Over Time.mp4 Shows thePointGrey camera perspective for a scene composedof posts separated by 5cm and starting at 27cm awayfrom the camera. Each frame shows the measured op-tical thickness.
2. Further Experimental DetailsSupplemental Fig. 1 shows the dimensions of the ‘E’
shape target. Our reconstructions in the main text demon-strate recovery with 1cm spatial resolution and 5cm axialresolution for targets at a range of 37-57cm.
3. Extension for Figure 3 from the Main TextSupplemental Figs. 2+3 provide results on top of Fig. 3
in the main text. Supplemental Fig. 2 shows results thatcorrespond to the left column of Fig. 3 in the main text
2cm1cm
1cm
1cm
1cm
1cm
1cm
Figure 1. Sketch of the E target. The actual target is composed offour identical E shapes separated by 5cm from one another.
(estimated KDE, Gamma distribution, signal distribution,and target distribution). Supplemental Fig. 3 shows resultsthat correspond to the right column of Fig. 3 in the main text(histogram and fitted model from Eq. 5). Both figures showresults for targets at five different depths (42, 47, 52, 57,62cm), and a background pixel. Both figures show resultsfor optical thicknesses (OT) of: 0, 1.0, 1.6, 1.9, 2.3, 2.8. Allplots are normalized to [0,1], and the title of each subplotindicates the total photons measured.
We note that when the object is closer, it can be recov-ered with more challenging OT. For example the closest ob-ject at 42cm is always correctly recovered, the next targetat 47cm is correctly recovered up to OT=2.3, and then failsfor OT=2.8, and the furthest target (62cm) fails as early asOT=1.6.
We also note that for the background, as soon as a littlefog is added, the number of counts rises very rapidly, andthen stays roughly the same for denser fog (the responseshape does change as more fog is added). The opposite pro-cess occurs for the targets – as more fog is added, we noteless counts.
Lastly, supplemental Fig. 3 shows that the suggestedmodel captures the measurements well at diverse levels offog (OT) and target depths.
Counts = 30x106
KDE fit Gamma fit Signal Target
Counts = 31x106 Counts = 27x106 Counts = 20x106 Counts = 14x106 Counts = 1x106
Counts = 13x106 Counts = 11x106 Counts = 7x106 Counts = 5x106 Counts = 3x106 Counts = 3x106
Counts = 7x106 Counts = 5x106 Counts = 4x106 Counts = 3x106 Counts = 2x106 Counts = 3x106
Counts = 6x106 Counts = 5x106 Counts = 4x106 Counts = 3x106 Counts = 2x106 Counts = 3x106
Counts = 4x106 Counts = 4x106 Counts = 3x106 Counts = 3x106 Counts = 2x106 Counts = 3x106
0 2Time [ns]
Counts = 3x106
0Time [ns]
Counts = 3x106
0Time [ns]
Counts = 3x106
0Time [ns]
Counts = 2x106
0Time [ns]
Counts = 2x106
0Time [ns]
Counts = 3x106
4 2 4 2 4 2 4 2 4 2 40
1
42cm 47cm 52cm 57cm 62cm BackgroundTarget At Distance
OT
= 0.
0O
T =
1.0
OT
= 1.
6O
T =
1.9
OT
= 2.
3O
T =
2.8
0
1
0
1
0
1
0
1
0
1
Figure 2. Extension to figure 3 from the main text. Recovered KDE and Gamma distributions (from the raw photon counts), estimatedsignals by subtracting the two, and estimated target distributions. Rows show different cases of fog concentrations (OT). Columns showdifferent target depths and a space with no target.
Histogram Model Fit
0 2Time [ns]
0Time [ns]
0Time [ns]
0Time [ns]
0Time [ns]
0Time [ns]
4 2 4 2 4 2 4 2 4 2 40
1
OT
= 0.
0O
T =
1.0
OT
= 1.
6O
T =
1.9
OT
= 2.
3O
T =
2.8
0
1
0
1
0
1
0
1
0
1
42cm 47cm 52cm 57cm 62cm BackgroundTarget At Distance
Counts = 30x106 Counts = 31x106 Counts = 27x106 Counts = 20x106 Counts = 14x106 Counts = 1x106
Counts = 13x106 Counts = 11x106 Counts = 7x106 Counts = 5x106 Counts = 3x106 Counts = 3x106
Counts = 7x106 Counts = 5x106 Counts = 4x106 Counts = 3x106 Counts = 2x106 Counts = 3x106
Counts = 6x106 Counts = 5x106 Counts = 4x106 Counts = 3x106 Counts = 2x106 Counts = 3x106
Counts = 4x106 Counts = 4x106 Counts = 3x106 Counts = 3x106 Counts = 2x106 Counts = 3x106
Counts = 3x106 Counts = 3x106 Counts = 3x106 Counts = 2x106 Counts = 2x106 Counts = 3x106
Figure 3. Extension to figure 3 from the main text. Histogram generated by the raw photon counts and the fitted model by Eq. 5. Rowsshow different cases of fog concentrations (OT). Columns show different target depths and a space with no target.
