Full-Spectrum and Infrared Cameras

Four things get called infrared in this hobby and they are not the same thing. Getting them straight is most of the work.

What removing the filter buys you

Every colour digital camera has an infrared-cut filter over the sensor, because silicon is sensitive well past the red end and without the filter your foliage goes pink and your skin tones fall apart. A full-spectrum conversion takes it out.

What you gain is real: in a dark building with an IR illuminator, the camera can see. That is a genuine and substantial advantage over trying to shoot at ISO 25600 in visible light, and it is why almost every serious night camera works this way.

What you also gain, and this is where the trouble is:

• Enormous lens flare. Coatings are designed for visible wavelengths. In IR they stop working, and every bright source in or near the frame throws ghosts, veils and coloured blobs. The characteristic purple and magenta artefacts in full-spectrum footage are flare and channel imbalance, not phenomena.
• Focus shift. Glass refracts IR differently from visible light. A lens focused visually is not focused in IR, so a converted camera without a recalibrated focus is soft, and soft edges are where shapes come from.
• Everything reflective becomes strange. Fabrics, paint, dust and skin have completely different IR reflectance from their visible appearance. Black clothes go pale, veins show through skin, and dust in the air lights up like snow.

That last one is worth its own paragraph. Orbs are dust. A particle close to the lens, outside the depth of field, lit from near the optical axis by an on-camera illuminator, images as a soft bright disc with the shape of the aperture. Move the illuminator off-axis and they mostly stop. This has been demonstrated repeatedly, is trivially reproducible in your own hallway with a torch and some flour, and remains the single most common thing submitted to us as evidence.

Thermal is a different instrument entirely

A thermal camera measures long-wave infrared emitted by objects because they are warm. It needs a microbolometer array, which is expensive, and germanium optics, because glass is opaque at those wavelengths. It cannot see through glass. It cannot see through walls. A window is a mirror to it.

It is genuinely useful for the ordinary half of an investigation: draughts, missing insulation, a cold spot that turns out to be an uninsulated lintel, damp, a hot cable run, a flue where a flue should not be. That is a real diagnostic capability and it is the reason to own one.

What it is not is a detector of anything. It shows you where the heat is. If there is a cold spot in the air it will not show you that either, because it measures surfaces, not air.

SLS, and what it is really doing

The original SLS cameras were built around a structured-light depth sensor from a games console — a projector throwing a known infrared dot pattern, a camera watching how the pattern deforms, and depth computed from the deformation. On top of that ran the console's skeletal tracking: a model trained to find human bodies and fit a stick figure to them.

That is the whole device. It is a body-tracking system, running in the dark, in a room it was not designed for.

The behaviour everyone knows about — a figure appearing where nobody is standing — is the tracker doing its job badly on ambiguous input, which is exactly what it does. A pose model asked to find people will find candidates in a coat on a hook, a chair back, a doorway, a radiator, a curtain fold. It is not malfunctioning. It has been asked a question and it is obliged to answer.

Our own array runs a real pose model on your camera and is built around keeping that distinction visible. It counts two numbers and never a total: tracked returns, which are landmark sets the model actually produced from your video, and reconstructions, which are figures the instrument drew. The register names a reconstruction as a reconstruction when it appears and again when it dissolves, every captured plate states how many of each are in it, and no code anywhere on this site adds the two together.

It also measures something real while it does it. The frame is divided into tiles, each scored on darkness, on frame-to-frame grain relative to its own signal, and on flatness, and the peak is printed as the resolution floor — an honest answer to "where in this picture can the array not tell what it is looking at". Point it at a lit wall and the figure falls. Point it into a dark corner and it climbs. Below the working gate the reconstruction rate is exactly zero, so in a well-lit room nothing will ever appear, however long you hold it there.

The one file in our archive where that measurement did something we could not explain is Hallow Gate, and it is about the camera rather than about a figure, which is the correct way round.