edit skeletondiffusion

Author: riccardomarin

_publications/skeletondiffusion/assets/method.png

@@ -52,12 +52,67 @@ <h2 class="title is-3">Abstract</h2>
     </div>
   </section>
 
+
+  <section class="section hero is-light">
+    <div class="container is-max-desktop">
+      <div class="columns is-centered has-text-centered">
+        <div class="column is-four-fifths">
+          <h2 class="title is-3">The Task</h2>
+          <div class="content has-text-justified">
+            <p>
+              In this work, we address the problem of predicting human motion based on observed past movements, known as Human Motion Prediction (HMP). Specifically, from a temporal sequence of human joint positions, we aim to forecast their evolution in subsequent frames.
+            </p>
+          </div>
+        </div>
+      </div>
+    </div>
+  </section>
+<img src="assets/method.png">
+  <section class="section hero is-light">
+    <div class="container is-max-desktop">
+      <div class="columns is-centered has-text-centered">
+        <div class="column is-four-fifths">
+          <h2 class="title is-3">Method Overview</h2>
+          <div class="content has-text-justified">
+            <p>
+We present SkeletonDiffusion, a latent diffusion model encoding this bias explicitly on both architecture and training.
+First, we consider the skeleton structure and joint categories throughout the entire network, and build our architecture end-to-end on top of Graph Convolutional Networks
+(GCNs). In contrast, existing SHMP approaches either ignore the skeleton’s graph structure or only
+leverage it at intermediate stage. 
+  <img src="assets/nonisodiff.png">
+
+  <br>
+Second, we replace the conventional isotropic
+Gaussian diffusion training with a novel nonisotropicformulation that accounts for joint relations directly in the
+generation process: the HMP problem is defined by the skeleton kinematic graph, and we exploit this knowledge to define a fixed non-diagonal noise covariance for the diffusion process
+
+            </p>
+          </div>
+        </div>
+      </div>
+    </div>
+  </section>
+
+<h2 class="title is-3">Results</h2>
   <section class="hero is-small">
     <div class="hero-body">
       <div class="container">
         <div class="columns is-centered has-text-centered">
           <div class="column is-four-fifths">
-            <h2 class="title is-3">Coming Soon</h2>
+            <div style="display: inline-block;text-align: center; margin:20px;">
+            <figure style="display: inline-block">
+            <img src="https://neural-icp.github.io/images/sample_2_input.gif" style="vertical-align: top;height:360;margin-bottom: -30px;">
+            <figcaption><b>Input</b></figcaption>
+            </figure>
+            <figure style="display: inline-block">
+            <img src="https://neural-icp.github.io/images/sample_2_smpl.gif" style="vertical-align:top;height:360;margin-bottom: -30px;">
+            <figcaption><b>Registration</b></figcaption>
+            </figure>
+            <figure style="display: inline-block">
+            <img src="https://neural-icp.github.io/images/sample_2_colors.gif" style="vertical-align:top;height:360;margin-bottom: -30px;">
+            <figcaption><b>Correspondences</b></figcaption>
+            </figure>
+            </div>
             <p class="content has-text-justified">
             </p>
           </div>