Datasets:
image imagewidth (px) 512 512 | scene stringclasses 3
values | container_bit_depth int64 8 16 | unique_levels int64 193 50.6k | kind stringclasses 3
values | min_value int64 0 0 | max_value int64 255 65.5k |
|---|---|---|---|---|---|---|
dome | 16 | 12,240 | true_16bit | 0 | 65,535 | |
dome | 16 | 253 | stretched_8bit_in_16bit | 0 | 65,535 | |
dome | 8 | 253 | plain_8bit | 0 | 255 | |
ramp | 16 | 50,604 | true_16bit | 0 | 65,535 | |
ramp | 16 | 256 | stretched_8bit_in_16bit | 0 | 65,535 | |
ramp | 8 | 256 | plain_8bit | 0 | 255 | |
medallion | 16 | 10,686 | true_16bit | 0 | 65,535 | |
medallion | 16 | 193 | stretched_8bit_in_16bit | 0 | 65,535 | |
medallion | 8 | 193 | plain_8bit | 0 | 255 |
Fake vs True 16-bit Depth PNGs
Nine small synthetic grayscale PNGs for testing whether a tool, script or pipeline can tell a genuine 16-bit depth map from an 8-bit image stretched into a 16-bit container.
A PNG header can say "16-bit" while the pixels hold only 256 distinct values. In laser engraving (for example LightBurn's 3D Sliced mode) and CNC relief work, those 256 values become 256 height steps, which can appear as stepping lines on the finished surface. These files are a ready-made positive and negative set for checking your own detection code.
This is a file-format test set. It is not a training set, not a benchmark of any depth-estimation model, and the scenes are procedural, not photographs.
The three scenes
All scenes are 512 × 512 and defined as a float height field between 0 and 1, built in generate.py:
dome: a smooth hemisphere. A continuous surface with no flat areas.ramp: a diagonal gradient with a gentle ripple added.medallion: a coin-like disc with a raised rim, a recessed field, a domed centre and a ring of bumps. The area outside the disc is flat at 0.
The three kinds of file
Each scene is saved three ways from the same height field f:
| Kind | How it is made | What to expect |
|---|---|---|
true_16bit |
round(f * 65535) written as a 16-bit grayscale PNG |
Header 16-bit, thousands of unique levels |
stretched_8bit_in_16bit |
round(f * 255) first, then multiplied by 257 and written as a 16-bit PNG |
Header 16-bit, at most 256 unique levels, all multiples of 257 |
plain_8bit |
round(f * 255) written as an 8-bit grayscale PNG |
Header 8-bit, at most 256 unique levels |
Measured values
metadata.csv is written by generate.py after the PNGs are saved. The bit depth is read from the PNG IHDR chunk and the unique levels are counted with numpy on the decoded pixels, so nothing in the table is assumed.
| File | Scene | Kind | PNG header bit depth | Unique levels | Min | Max |
|---|---|---|---|---|---|---|
images/dome-true-16bit.png |
dome | true_16bit | 16 | 12,240 | 0 | 65,535 |
images/dome-stretched-8bit-in-16bit.png |
dome | stretched_8bit_in_16bit | 16 | 253 | 0 | 65,535 |
images/dome-plain-8bit.png |
dome | plain_8bit | 8 | 253 | 0 | 255 |
images/ramp-true-16bit.png |
ramp | true_16bit | 16 | 50,604 | 0 | 65,535 |
images/ramp-stretched-8bit-in-16bit.png |
ramp | stretched_8bit_in_16bit | 16 | 256 | 0 | 65,535 |
images/ramp-plain-8bit.png |
ramp | plain_8bit | 8 | 256 | 0 | 255 |
images/medallion-true-16bit.png |
medallion | true_16bit | 16 | 10,686 | 0 | 65,535 |
images/medallion-stretched-8bit-in-16bit.png |
medallion | stretched_8bit_in_16bit | 16 | 193 | 0 | 65,535 |
images/medallion-plain-8bit.png |
medallion | plain_8bit | 8 | 193 | 0 | 255 |
The 8-bit kinds show fewer than 256 levels for dome (253) and medallion (193) because those scenes do not visit every possible level. The ramp uses all 256. None of the true 16-bit files uses all 65,536 values either; they use 10,686 to 50,604. A low count on its own is not proof of an 8-bit origin, so read it together with the header and the multiples-of-257 pattern.
I also ran ImageMagick 7.1.2 (Q16 build) as identify -format "%z-bit | %k unique levels\n" file.png on all nine files. Its bit depth and unique-level numbers matched the table for every file.
Files
images/: the nine PNGsmetadata.csv:file_name,scene,container_bit_depth,unique_levels,kind, plusmin_valueandmax_valuegenerate.py: regenerates everything (requiresnumpyandPillow):python generate.py
How to check a file yourself
Read the header bit depth, then count the unique values:
import numpy as np
from PIL import Image
path = "images/dome-stretched-8bit-in-16bit.png"
with open(path, "rb") as fh:
bit_depth = fh.read(26)[24] # IHDR bit depth
levels = np.unique(np.array(Image.open(path)))
print(bit_depth, levels.size, bool(np.all(levels % 257 == 0))) # 16 253 True
To check any PNG in your browser instead, use the 16-bit depth map checker, which compares the header bit depth against the measured unique level count.
Limitations
- Synthetic scenes only, one resolution (512 × 512), grayscale only, no noise and no photographic content.
- Real exporters stretch differently (bit shifts, gamma, partial ranges).
v * 257is the common case, not the only one. - Only the common kind of fake is covered. A file that has been blurred, dithered or noise-added after stretching will have more unique values than 256 without being a true 16-bit capture.
- The Hugging Face dataset viewer may render 16-bit PNGs as 8-bit previews. Download the files and measure them rather than trusting the preview.
License
CC BY 4.0. Attribution: Depth-Map.com. The images are generated by generate.py and contain no third-party content.
- Downloads last month
- -