We provide several datasets that we acquire with a computational hyperspectral imager.
License: The SPIHIM datasets are distributed under the Creative Commons Attribution 4.0 International license (CC-BY 4.0)
Reference: Please reference this work
- A Lorente Mur, B Montcel, F Peyrin, N Ducros. Deep neural networks for single-pixel compressive video reconstruction. SPIE Photonics Europe, Proc. vol. 11351, Unconventional Optical Imaging II, pp.27, Apr 2020, France. ⟨hal-02547800⟩
Contact: [email protected], CREATIS Laboratory, University of Lyon, France.
The set-up is depicted below and in detail [ALM2020]. The telecentric lens (TL; Edmund Optics 62901) is positioned such that its image side projects the image of the sample (S) onto the digital micro-mirror device (DMD; vialux V-7001), which is positioned at the object side of the lens. The object is transparent and is illuminated by a LED lamp (L; Thorlabs LIUCWHA/M00441662). The DMD can implement different light patterns by reflection of the incident light onto a relay lens (RL), which projects the light into an optical fiber (OF; Thorlabs FT1500UMT 0.39NA). This optical fiber is connected to a compact spectrometer (SM; BWTek examplar BRC115P-V-ST1). A filter wheel (FW) containing neutral optical densities is placed behind the lamp to reduce the light flux
The setup acquires M = α HF where F in R n x λ represents the sample hypercube, H in R m x n the measurement matrix, and α is a multiplicative factor that depends on and the optical density. The measurement matrix contains the patterns that are uploaded onto the DMD. Here, m represents the number of patterns, n the number of pixels of the patterns, and λ the number of spectral bins.
The multiplicative factor is given by α = φ 10-OD Δt (in photons), where φ (in photons/s) represents the given light flux, OD is neutral optical density and Δt is the integration time.
We acquire Hadamard patterns that we split into positive and negative parts, which are concatenated in the measurement matrix. Precisely, Hm contains the positive parts and Hm+1 the negative parts, such that Hm - Hm+1 is a Hadamard pattern.
We sequentially upload onto the DMD all of the m = 2 x 4096 Hadamard (split) patterns of dimension n = 64 x 64 pixels. The patterns in PNG format can be downloaded here.
We acquire different datasets for the same object by selecting different neutral densities OD and different integration times Δt.
The following datasets are provided
- 04-Feb-2020 session (description). Download zip.
- 11-Jun-2020 session (description). Download zip.
- 01-Jul-2020 session (description). Download zip
- 18-Nov-2020 session (description). Download zip
- All datasets compatible with the SPYRIT deep reconstruction networks (e.g., Comp-Net). Download zip
We provide the description of each measurement session in the sections below.
Based on SPIRiT, we provide matlab scripts that
- read, reconstruct and plot the datasets (see
./scripts/read_*_.m) - build the forward operator H that maps the image of the sample the onto the measured Hadamard coefficients (see
./scripts/build_forward_operator.m)
The (full) Hadamard matrix H in R n x n can be downloaded here. It applies to a measurement vector in R n, where the missing coefficients have been field with zeros. To compute the (reduced) measurement matrix H in R m x n, see ./scripts/build_forward_operator.m.
The acquisition is such that the patterns with maximum variance are acquired first. We provide here the covariance matrix used to defined the acquisition order ; it was computed on the STL-10 image dataset.
We acquired four samples:
- A star sector target printed on a paper sheet in black an white
- A star sector target printed on a paper sheet printed in color according to a Hue color wheel
- A paper sheet with no printing
- No object, i.e., the illumination LED lamp directly
| Filename | M | Δt (ms) | Comment |
|---|---|---|---|
| SiemensBW1a_raw.mat | 408 | 4 | black and white star sector |
| SiemensColor8a_raw.mat | 408 | 32 | color star sector |
| SiemensColor8b_raw.mat | 1228 | 32 | color star sector |
| mydata_raw.mat | 512 | 4 | |
| paperSheet1b_raw.mat | 1228 | 4 | paper sheet |
| paperSheet2b_raw.mat | 1228 | 8 | paper sheet |
| paperSheet4b_raw.mat | 1228 | 16 | paper sheet |
| siemensStar1a_raw.mat | 408 | 4 | black and white star sector |
| siemensStar1b_raw.mat | 1228 | 4 | black and white star sector |
| siemensStar2a_raw.mat | 408 | 8 | black and white star sector |
| siemensStar2b_raw.mat | 1228 | 8 | black and white star sector |
| siemensStar4a_raw.mat | 408 | 16 | black and white star sector |
| siemensStar4b_raw.mat | 1228 | 16 | black and white star sector |
| siemensStar8a_raw.mat | 408 | 32 | black and white star sector |
| siemensStar8b_raw.mat | 1228 | 32 | black and white star sector |
| siemensStarColor1b_raw.mat | 1228 | 4 | color star sector |
| siemensStarColor2b_raw.mat | 1228 | 8 | color star sector |
| siemensStarColor4b_raw.mat | 1228 | 16 | color star sector |
| siemensStarColor4c_raw.mat | 2456 | 16 | color star sector |
| siemensStarColor8d_raw.mat | 4096 | 32 | color star sector |
| whiteLED1a_raw.mat | 408 | 4 | no object |
First, we acquired the LED source directly with no sample
| Filename | Δt (ms) | OD (-) | Comment |
|---|---|---|---|
| noObjectD_1_0.0_raw.mat | 4 | 0.0 | Saturate |
| noObjectD_1_0.3_01_raw.mat | 4 | 0.3 | Repeat same measurement 4x to simulate longer Δt while avoiding saturation |
| noObjectD_1_0.3_02_raw.mat | 4 | 0.3 | |
| noObjectD_1_0.3_03_raw.mat | 4 | 0.3 | |
| noObjectD_1_0.3_04_raw.mat | 4 | 0.3 | |
| noObjectD_1_0.3_raw.mat | 4 | 0.3 | |
| noObjectD_1_0.6_raw.mat | 4 | 0.6 | |
| noObjectD_1_1.0_raw.mat | 4 | 1.0 | |
| noObjectD_1_1.3_raw.mat | 4 | 1.3 |
Second, we acquired the star sector target
| Filename | Δt (ms) | OD (-) | Comment |
|---|---|---|---|
| starSectorD_2_0.0_01_raw.mat | 8 | 0.0 | Repeat same measurement 4x to simulate longer Δt while avoiding saturation |
| starSectorD_2_0.0_02_raw.mat | 8 | 0.0 | |
| starSectorD_2_0.0_03_raw.mat | 8 | 0.0 | |
| starSectorD_2_0.0_04_raw.mat | 8 | 0.0 | |
| starSectorD_2_0.0_05_raw.mat | 8 | 0.0 | |
| starSectorD_2_0.0_06_raw.mat | 8 | 0.0 | |
| starSectorD_2_0.0_07_raw.mat | 8 | 0.0 | |
| starSectorD_2_0.0_08_raw.mat | 8 | 0.0 | |
| starSectorD_2_0.0_09_raw.mat | 8 | 0.0 | |
| starSectorD_2_0.0_10_raw.mat | 8 | 0.0 | |
| starSectorD_2_0.0_raw.mat | 8 | 0.0 | |
| starSectorD_2_0.3_raw.mat | 8 | 0.3 | |
| starSectorD_2_0.6_raw.mat | 8 | 0.6 | |
| starSectorD_2_1.0_raw.mat | 8 | 1.0 | |
| starSectorD_2_1.3_raw.mat | 8 | 1.3 |
Third, we acquire the star sector target that was off-centered
| Filename | Δt (ms) | OD (-) | Comment |
|---|---|---|---|
| R1L1S2P_1_0.0_raw.mat | 4 | 0.0 | Shows the 'R1L1S2P' writing |
| R1L1S2P_1_0.3_raw.mat | 4 | 0.3 | |
| R1L1S2P_1_0.6_raw.mat | 4 | 0.6 | |
| R1L1S2P_1_1.0_raw.mat | 4 | 1.0 | |
| R1L1S2P_1_1.3_raw.mat | 4 | 1.3 | |
| Thorlabs_1_0.0_raw.mat | 4 | 0.0 | Shows the 'THORLABS' writing |
| Thorlabs_1_0.3_raw.mat | 4 | 0.3 | |
| Thorlabs_1_0.6_raw.mat | 4 | 0.6 |
Finally, we acquire images from the STL10 databaset
| Filename | Δt (ms) | OD (-) | Comment |
|---|---|---|---|
| stl10_01_32ms_cp_raw.mat | 32 | 0.0 | horse displayed on my cell phone (cp) |
| stl10_01_32ms_ps_raw.mat | 32 | 0.0 | horse printed on a paper sheet (ps) |
| stl10_01_64ms_ps_raw.mat | 64 | 0.0 | horse printed on a paper sheet (ps) |
| stl10_08_32ms_ps_raw.mat | 32 | 0.0 | bird printed on a paper sheet (ps) |
First, we acquired the LED source directly with no sample. Note that the noObjectF1 series and the noObjectF2 series refer to two different positions of the LED lamp.
| Filename | Δt (ms) | OD (-) | Comment |
|---|---|---|---|
| noObjectF1_1_0.3_01_raw.mat | 4 | 0.3 | Repeat same measurement 15x to simulate longer Δt while avoiding saturation |
| noObjectF1_1_0.3_02_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_03_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_04_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_05_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_06_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_07_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_08_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_09_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_10_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_11_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_12_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_13_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_14_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_15_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.3_raw.mat | 4 | 0.3 | |
| noObjectF1_1_0.6_raw.mat | 4 | 0.6 | |
| noObjectF2_1_0.3_01_raw.mat | 4 | 0.3 | Not the same LED lamp position as in the noObjectF1 series. We repeat same measurement 9x to simulate longer Δt while avoiding saturation |
| noObjectF2_1_0.3_02_raw.mat | 4 | 0.3 | |
| noObjectF2_1_0.3_03_raw.mat | 4 | 0.3 | |
| noObjectF2_1_0.3_04_raw.mat | 4 | 0.3 | |
| noObjectF2_1_0.3_05_raw.mat | 4 | 0.3 | |
| noObjectF2_1_0.3_06_raw.mat | 4 | 0.3 | |
| noObjectF2_1_0.3_07_raw.mat | 4 | 0.3 | |
| noObjectF2_1_0.3_08_raw.mat | 4 | 0.3 | |
| noObjectF2_1_0.3_09_raw.mat | 4 | 0.3 | |
| noObjectF2_1_0.3_raw.mat | 4 | 0.3 | Not the same LED lamp position as in the noObjectF1 series |
| noObjectF2_1_0.6_raw.mat | 4 | 0.6 | |
| noObjectF2_1_1.0_raw.mat | 4 | 1.0 | |
| noObjectF2_1_1.3_raw.mat | 4 | 1.3 |
| Filename | Δt (ms) | OD (-) | Comment |
|---|---|---|---|
| stl10_03_1_0.0_raw.mat | |||
| stl10_05_1.5_0.0_01_raw.mat | |||
| stl10_05_1.5_0.0_02_raw.mat | |||
| stl10_05_1.5_0.0_03_raw.mat | |||
| stl10_05_1.5_0.0_04_raw.mat | |||
| stl10_05_1.5_0.0_05_raw.mat | |||
| stl10_05_1.5_0.0_06_raw.mat | |||
| stl10_05_1_0.0_raw.mat | |||
| stl10_05_1_0.3_raw.mat | |||
| stl10_05_1_0.6_raw.mat |
Contrary to previous datasets, the DMD frequency is set to 1 kHz and the DMD patterns are 4-bit (see par.DMD_fr and par.bitplane).
| Filename | M | Δt (ms) | Comment |
|---|---|---|---|
| noObject_01ms_raw.mat | 8192 | 1 | |
| noObject_10ms_raw.mat | 8192 | 10 | |
| starSector_01ms_raw.mat | 8192 | 1 | Starsector in the image plane |
| starSector_variableFilter_C_10ms_raw.mat | 8192 | 10 | Starsector in the image plane, filter between lamp and image plane |
| variableFilter_01ms_raw.mat | 8192 | 1 | Filter in the image plane |
| variableFilter_10ms_raw.mat | 8192 | 10 | Filter in the image plane |
| variableFilter_B_01ms_raw.mat | 8192 | 1 | Filter in the image plane |
| variableFilter_B_10ms_raw.mat | 8192 | 10 | Filter in the image plane |
| variableFilter_C_10ms_raw.mat | 8192 | 10 | Filter in the image plane ` |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent