Synthetic Avatar Dataset for Lateral-Shift Estimation and Preprocessed Segmentation Data

<b>Summary</b>This dataset contains synthetically generated human avatar images and corresponding labels designed to support machine-learning–based <b>lateral-shift estimation</b> in the context of cervical posture analysis. The dataset includes the raw rendered images, JSON-based ground-truth labels derived from the underlying avatar rig, segmentation-based preprocessed arrays, and the trained model weights used for lateral-shift prediction.All data are fully synthetic, contain no real individuals, and were generated via a scripted Blender pipeline. Ground-truth labels reflect precise rig parameters and provide continuous values without human rating noise.<b>Repository Structure</b><pre><pre>/<br>├── labels/<br>│ └── *.json<br>├── renders/<br>│ └── *.png / *.jpg<br>├── segmentation_maps/<br>│ ├── X.npy<br>│ └── y.npy<br>└── model_weights/<br> └── lateral_shift_model.ckpt<br></pre></pre>Each component is described below.<b>1. </b><code><strong>renders/</strong></code><b> – Synthetic Avatar Images</b>This folder contains the rendered RGB images of synthetic human avatars.<br>Characteristics:About 16000 Frontal-view imagesSynthetic backgrounds, lighting, clothing, and morphology randomized to increase diversityImages produced in Blender using the MPFB add-on for human model generationThese rendered images serve as the <b>input</b> for the lateral-shift estimation model.<b>2. </b><code><strong>labels/</strong></code><b> – JSON Ground-Truth Labels</b>Each image has a corresponding JSON file in <code>labels/</code> containing the precise parameters exported from the avatar rig.<br>An example entry (user-provided) looks as follows:labels.0415_123132_730584<pre><pre>{<br> "lateral_shift": 3.55,<br> "yaw_head": -6.08,<br> "yaw_neck": 2.48,<br> "roll_head": -3.55,<br> "roll_neck": 9.32,<br> "pitch_head": 0,<br> "pitch_neck": 0.056,<br> "shoulder_elevation": true<br>}<br></pre></pre>About these labels<code><strong>lateral_shift</strong></code><br>Main target variable. Continuous, rig-derived value indicating lateral translation between head and neck.<b>Additional rig parameters (</b><code><strong>yaw_*</strong></code><b>, </b><code><strong>roll_*</strong></code><b>, </b><code><strong>pitch_*</strong></code><b>)</b><br>These describe the underlying avatar pose and were part of the physical rig used to generate variation.<br>They are included as metadata for transparency and may be useful for future downstream experiments, but are <b>not</b> used in the lateral-shift model included here unless the user actively incorporates them.<code><strong>shoulder_elevation</strong></code><br>Boolean indicator reflecting whether the shoulder was elevated in the avatar’s pose.<br>Included because shoulder asymmetry can influence overall posture geometry.These labels are noise-free and derived directly from the 3D rig.<b>3. </b><code><strong>segmentation_maps/</strong></code><b> – Preprocessed Arrays</b>This directory contains two NumPy arrays:<code><strong>X.npy</strong></code><br>Preprocessed avatar images after the full pipeline described in the manuscript:semantic segmentation (via Meta’s Sapiens model)cropping using head–neck region boundariescentering based on mass-center alignmentaspect-ratio normalization<br>This corresponds to the input actually fed into the lateral-shift model.<code><strong>y.npy</strong></code><br>Array of continuous lateral-shift ground-truth values aligned with <code>X.npy</code>.These arrays are meant for quick loading into training pipelines without having to re-compute segmentation and cropping.<b>4. </b><code><strong>model_weights/</strong></code><b> – Trained Lateral-Shift Regression Model</b>This folder contains the final trained model weights for <b>EfficientNet-B0</b> (or your architecture) used for lateral-shift prediction.Included:Model weight file (<code>.ckpt</code>)