Skip to content

PQ/PU21 Encoding Does Not Help Narrow-Range LWIR Data

This is a struggle / negative-knowledge article documenting why perceptual quantization (PQ) and PU21 encoding — despite impressive headline gains of 2-9 dB on HDR data — provide negligible benefit for our LWIR super-resolution task.

The Hypothesis

SIGGRAPH Asia 2024 showed that PQ and PU21 perceptual encoding yields 2-9 dB improvement on HDR super-resolution data. The idea was appealing: if simply re-encoding our input data could produce multi-dB gains, this would be an easy win before touching the architecture or training pipeline.

The Failure Mode

After reading the actual paper, the gains turned out to be specific to visible-light HDR data spanning 6 orders of magnitude (approximately 0.005-4000 nits). This extreme dynamic range means that naive linear encoding wastes most of its precision on bright values while under-representing the perceptually important dark regions. PQ and PU21 redistribute precision according to human visual perception, which produces large gains when the input range is vast.

Our LWIR data does not have this property.

The Root Cause

Our sensor data spans approximately 30,000-34,000 DN — a ratio of barely 1.13:1. This is extremely low dynamic range. The entire signal of interest occupies a narrow band within the 16-bit range.

Three specific reasons PQ/PU21 cannot help:

  1. No precision redistribution opportunity. PQ/PU21 gains come from moving precision from over-represented bright regions to under-represented dark regions. With a 1.13:1 range, there is no meaningful imbalance to correct. The signal is already concentrated in a narrow band.

  2. PQ/PU21 require absolute luminance units. These encodings are defined in terms of physical luminance (nits). Our sensor produces uncalibrated digital numbers that do not correspond to absolute luminance. Applying PQ to arbitrary DNs is not meaningful — the perceptual curve is calibrated to human vision of visible light, not thermal radiation.

  3. Existing normalization already solves the problem. Our per-sequence normalization (subtract mean, divide by standard deviation) maps the narrow DN range into zero-centered, unit-variance space. This is already an efficient encoding for the actual information content. There is nothing for PQ/PU21 to improve upon.

Realistic expected gain for our case: negligible to ~0.5 dB at most.

The Anti-Pattern

Do not assume HDR techniques apply to narrow-range sensor data. The "HDR" label and impressive dB gains from perceptual encoding papers are specific to data with large dynamic range (multiple orders of magnitude). Narrow-range data — whether from LWIR sensors, specific industrial sensors, or any instrument measuring a small signal window within a larger bit depth — does not benefit from perceptual redistribution. Always check the dynamic range ratio of your actual data before adopting HDR techniques.

What We Use Instead

Per-sequence statistical normalization (mu/sigma) remains the correct approach:

x_normalized = (x - mean(sequence)) / std(sequence)

This centers the data at zero with unit variance, making efficient use of the float32 range for the actual signal content. It is simple, fast, and well-matched to our data characteristics.

Decision

Keep current normalization. PQ/PU21 encoding is not a priority for LWIR MFSR.

This finding was part of the broader architecture and training review documented in Existing Data Limitations and the brainstorming notes. The time saved by not pursuing this dead end was redirected toward more impactful changes: Data Collection v2 and Evaluation Strategy.