HandNet Supplementary material
BRITISH MACHINE VISION CONFERENCE · 2015

Rule of Thumb: Deep Derotation
for Improved Fingertip Detection

Aaron Wetzler · Ron Slossberg · Ron Kimmel

Technion — Israel Institute of Technology, Haifa, Israel

Paper ↗ Original code ↗ Dataset ↗ Citation

Overview

HandNet pairs real depth images with magnetic measurements of the position and orientation of five fingertips and the palm. The accompanying paper introduces DeROT: estimate global hand orientation, normalize in-plane rotation, and detect fingertips in the normalized image. This supplementary viewer exposes the released data and the geometry of ground-truth normalization.

The viewer uses stored samples from the original archives. It does not run a trained network or report new benchmark results.

Released frames
214,971
Depth resolution
320 × 240
Tracked locations
5 fingertips + palm
Coordinates
Camera XYZ · mm
Data license
CC BY 4.0
1

Dataset visualization

Loading stored samples

Select a released split and variant, inspect a frame, or compare its matched original/transformed counterpart. All displayed images and measurements originate in the source MAT files.

— / —
Fig. 1 · Calibrated depth point cloudLOADING
Drag: orbit · Scroll: zoom · Shift + drag: pan
X Y ZSensor orientation axes
Fig. 2 · Source camera image320 × 240
NEAR
FAR
Source depth image with annotation locations.
ALL SIX SENSORS

Select a sensor to inspect its coordinates.

SOURCE FRAMEFrame provenanceSource archive ↗

Figures 1–2. The surface is reconstructed from calibrated depth and displayed at a two-pixel stride within the source foreground mask. Spheres mark measured sensor centers; 2D markers use source heatmap peaks. The six tracked locations are not a complete hand-joint skeleton. Color and contrast normalization are for display.

Sensor measurements

Recorded camera frame · mm
SensorXYZuv
Loading measurements…

XYZ: recorded magnetic sensor centers. u, v: zero-based source annotation heatmap peaks, before any interactive rotation.

Frame provenance

Original archive member
Source
Loading…
Calibration and display conventions

The native camera coordinate system is +X right, +Y down, +Z away, with positions in millimetres and zero-based image coordinates. Depth reconstruction uses the original camera calibration:

X = (u − cx) Z / fx     Y = (v − cy) Z / fy

fₓ = 240.30773969018747   fᵧ = 240.07488537862218
cₓ = 178.29264757987775   cᵧ = 117.9581118815394

The 3D display flips Y and Z together to preserve handedness. Numeric measurements retain the source coordinate system. Sensor axes use the columns of each recorded 3 × 3 orientation matrix. Heatmap peaks can differ from projected sensor centers, particularly in transformed archives.

2

Ground-truth orientation normalization

Method in the paper ↗

From orientation regression to a canonical image

The paper estimates a global hand orientation and uses it to choose an in-plane rotation before fingertip detection. Removing rotational variability is intended to reduce the appearance space that the detector must learn.

  1. Measure or estimate orientation. HandNet provides magnetic orientation annotations; the published DeROT method learns to predict orientation from depth.
  2. Normalize image rotation. An orientation-dependent rule chooses a canonical in-plane view, with the thumb base on a consistent side.
  3. Detect and transform back. Fingertip detections in the normalized view are mapped to the original camera frame.
Scope of this demonstration.

The controls use the recorded palm orientation as ground truth. They demonstrate the geometric normalization convention from the original code. No orientation network or fingertip detector is executed.

Reference implementation: derotate.m and rotateData.m.

Fig. 3 · In-plane rotation

Inspect selected frame ↑
(a) Recorded depth
(b) Rotated depth
−180°+180°

This geometric demonstration uses measured palm orientation. It does not run the paper’s learned network.

Figure 3. Rotation is applied around the calibrated optical center in 2D and the camera Z axis in 3D. This changes the displayed geometry; it is not a new recorded capture. Numeric sensor tables above remain the recorded source measurements.

3

Data & reproducibility

Dataset repository ↗

Original release archives

The repository preserves all seven original distribution archives: standard training, test, and validation splits; ground-truth-derotated training and validation; and rotation-perturbed test and validation data.

Standard splitReleased MAT frames
Training202,198
Test10,000
Validation2,773
Total214,971

Data: CC BY 4.0. New viewer code: MIT. The historical MATLAB source retains its original terms.

Provenance and display subset

Stored examples include their source archive member, session and frame indices, and MAT SHA-256. The local sample manifest records calibration, units, sampling and display preprocessing. The archive manifest records original byte sizes and checksums.

Historical frame counts

The paper reports 212,928 unique hand poses. The released standard archives contain 214,971 MAT frames. This viewer reports the distributed archive counts and does not claim to reconcile the discrepancy.

Citation

Aaron Wetzler, Ron Slossberg, and Ron Kimmel. Rule of Thumb: Deep Derotation for Improved Fingertip Detection. British Machine Vision Conference, 2015.

@inproceedings{wetzler2015rule,
  title={Rule of Thumb: Deep Derotation
    for Improved Fingertip Detection},
  author={Wetzler, Aaron and Slossberg, Ron
    and Kimmel, Ron},
  booktitle={British Machine Vision Conference},
  year={2015}
}