**Transient Image Files**

The TI format stores transient images ${\boldsymbol{I}}(u,v,\tau)$ in a binary file with .ti extension. It is specifically tailored to the context of NLoS reconstruction, as it contains not only the measured light but also geometrical information about the scene that can be used during the reconstruction.

!!! Tip
Implementations of this file format for multiple languages can be found in the [toolbox](/toolbox).

Basics
======

Each file consists of four Blocks:

Block name           | Description
---------------------|------------
Pixel data           | Linear raw array filled with transient histograms
Pixel interpretation | Geometric arrangement of transient histograms
Image properties     | Flexible meta information as JSON encoded string

The ti format is intended to be simple but versatile. Actual image data is stored in raw arrays and can be read with a few lines of code in most languages without the need of external libraries. Additional features like compression that are found in other formats such as HDF5 were not included on purpose to keep the format simple.

All transient histograms are **unwarped**, meaning that the time-of-flight of the first (light source to reflector) and forth (reflector to detector) path segment are already removed. In the image properties block additional information about the light source and detector position can be given, so that the full path length can be reconstructed if needed.

Furthermore, all data is stored in little-endian mode and no implicit padding bytes are used.

Bin mapping
---------------

The current file format version assumes equidistant sampling in the time domain but supports both regular and flexible sampling in the spatial domains (through the use of different pixel modes).

The continuous signal is discretized into time bins, where each bin is denoted by its center time:

![Figure [BinMapping]: Bin mapping](/images/TemporalSampling.svg)

!!! Warning
This image is a simplification. In reality, the signal is supposed to be filtered properly and thus peaks are distributed among neighbouring cells.

===========

The header always has a constant size (28 bytes) and contains basic information about the image file.

Size     | Type             | Name
---------|------------------|---------------
4 byte   | signed int8[4] | magicNumber = 'TI04' = [84, 73, 48, 52]
4 byte   | unsigned int32 | pixelMode
4 byte   | unsigned int32 | numPixels
4 byte   | unsigned int32 | numBins
4 byte   | float32        | tMin
4 byte   | float32        | tDelta
4 byte   | unsigned int32 | pixelInterpretationBlockSize

magicNumber
: Identifies this file as a transient images. Also contains the file format version (currently 04).

pixelMode
: Block type used to store pixel interpretations

numPixels
: Total number of pixels in the image

numBins
: Amount of float values that are stored for each pixel

tMin
: Path length corresponding to the zeroths bin

tDelta
: Thickness of each time bin in path length

pixelInterpretationBlockSize
: Size of the pixel interpretation block in bytes. In Mode 10 this value is 68.

!!! Tip
Implementations should first only read the magicNumber (first four bytes) to check file version, as the rest of the file header might change in future releases.

The header helps to identify the format version and the size of the following blocks. The transient data block size is not explicitly stored as it can be computed easily as numPixels*numBins*sizeof(float32). The image properties block occupies the rest of the file.

In Figure [BinMapping], tMin corresponds to $t_0$ and tDelta to $t_1 - t_0$. The image does not contain any information about light arriving before tMin - tDelta/2.

Pixel data
==========
Transient pixels (a histogram of the amount of light that the pixel received in each time bin) are stored as one big, linear float array:

Size                         | Type                          | Name
-----------------------------|-------------------------------|---------------
numPixels*numBins*4 byte   | float32[numBins][numPixels] | pixelData

The values are stored pixel-major and not bin-major. Thus, all bin values for a single pixel are directly next to each other in memory. To access the fifth temporal bin in the tenth pixel, the following C++-Code should be used: pixelData[numBins*10 + 5].

Pixel interpretation
====================
The pixel interpretation block contains information about the observed and illuminated point on the reflector for each pixel (i.e. transient histogram). In principle arbitrary combinations are possible, but most of the time only a single positions is illuminated (or observed), and a regular grid of positions is observed (or illuminated). In these cases, more efficient storage for the sample positions is possible, leading to the different pixel interpretation modes.

!!! Tip:
All files of the benchmark are stored in Mode 10.

In the future, more modes might be added. However, it is not advantageous to have a dedicated mode for every special case, as this complicates the format, thus we try to keep the amount limited.

### Mode 0
In this most general case, each pixel has a observation and illumination position associated.

Size      | Type             | Name
----------|------------------|-----------
12  byte  | float32[3]     | laserOrigin
12  byte  | float32[3]     | laserNormal
12  byte  | float32[3]     | cameraOrigin
12  byte  | float32[3]     | cameraNormal

This block is repeated for each pixel. Thus, the total size is 48 byte * numPixels.

### Mode 10
The scene is illuminated from a single point on the reflector (laserPosition) and the reflector is observed in a rectilinear grid (i.e. tessellation by non-congruent parallelograms). The grid is the result of a Cartesian grid being projected onto the reflector via a homography. This homography is defined by the coordinates of the four corner pixel on the reflector (topLeft, topRight, bottomLeft, bottomRight). This mode is currently used in all benchmark scenes.

Size      | Type             | Name
----------|------------------|-----------
4  byte   | unsigned int32 | uResolution
4  byte   | unsigned int32 | vResolution
12 byte   | float32[3]     | topLeft
12 byte   | float32[3]     | topRight
12 byte   | float32[3]     | bottomLeft
12 byte   | float32[3]     | bottomRight
12 byte   | float32[3]     | laserPosition

It is assumed that the wall is planar (topLeft, topRight, bottomLeft, bottomRight and laserPosition all lie in the same plane). Thus, the reflector normal is the same for the illumination point and all observed positions and can be computed from the given points on the plane.

The transient pixels are stored in a row major order. With $u$ as the dimension left-right and $v$ as the dimension top-bottom, a value $I(u, v, \tau)$ is stored at data[(v*uResolution+u)*numBins+t]

!!! Tip
bottomRight == topRight+bottomLeft-topLeft can be used to checked, whether a homography is actually needed.

### Mode 20
This mode is the reciprocal case of Mode 1, where the role of the camera and laser are flipped: The wall is illuminated at multiple points while only a single one is observed. Due to reciprocity of light transport, this mode behaves exactly as Mode 10 and is only included for semantic reasons. Only the laserPosition field should be interpreted as cameraPosition (e.g. describing, on which point on the wall a single pixel detector is pointed).

Image properties
================
A text-based, UTF-8 encoded JSON string containing properties of the image. This block follows directly after the pixel interpretation block and spans the whole rest of the file. Thus, the size is not explicitly stated in the previous parts of the file, allowing it to be edited in a binary-compatible text editor (like Notepad++) without adapting any binary fields. For readability in text editors, it should start with a couple newline characters.

This block is intended to hold arbitrary meta-data of the image. The JSON encoding makes it easy to store any type of information in any order while still being easy to read wherever a JSON implementation is available. Images from different sources may use different JSON schemes. Users are encourage to create their own (with a unique File:MetadataVersion ID), they should however be inspired by existing usages to increase interoperability. At this point we refrain from providing a complete standard for schemes due to the richness of possible setups.

The files provided in this benchmark use the following fields:

~~~~~~~ json
{
"File":{
"RecordingTime": "2018-09-03"
},

"Challenge":{
"Name": "Geometry reconstruction",
}
}
~~~~~~~

An example of a different (fictive) usage:

~~~~~~~ json
{
"File":{
"RecordingTime": "2018-09-03",
"Setup": "Default setup"
},

"Camera": {
"Name": "LightTrap SciCapture 18mx",
"Position": [0.25, 1.37, -10],
"IntegrationTime in ms": 200,
"Lens": {
"Name": "Lensware DX TelePro",
"FocalLength in mm": 15,
"Filter": "FilterMax 650"
}
},

"LightSource": {
"Name": "UltraPointer Brutzelfleck 1500",
"Position": [0.43, 1.22, -10],
"Power in W": 1.5
},

"Experiment": {
"Name": "Bounce analysis",
"Executing technician": "Dr. Susi Sauerbraten",
"Laboratory": "Room A38"
}
}
~~~~~~~