force recovery: methods, evaluation, debug log

index.html

@@ -63,7 +63,7 @@ map, then calibrated on presses of known load.</p>
 ground-truth force</span></div>
 <div class="card"><b>5</b><span>force-labelled datasets, one
 protocol</span></div>
-<div class="card"><b>0.868</b><span>agreement between two
+<div class="card"><b>0.877</b><span>agreement between two
 estimators sharing no calibration</span></div>
 <div class="card"><b>0</b><span>frames of our own rig in any
 calibration</span></div>

method.html

@@ -70,6 +70,34 @@ on one direction and <code>(gx,&nbsp;gy)</code> is a 3×2 least-squares solve.
 No table, no sphere presses. This is the GelSight Wedge driver's approach. It
 recovers shape but not scale.</p>
 
+<h2>Stage 1 scored on its own — no force labels</h2>
+<p>Force estimation is image→depth then depth→newtons, and a ρ only ever scores
+the pair: a geometrically wrong depth that is monotone in contact size still
+ranks force well. Depth has no ground truth, so stage 1 is judged by eye on the
+panels below and by physical checks that need no labels.</p>
+
+<div class='tablewrap'><table><thead><tr><th>dataset</th><th>n</th><th>flat-gel leak, LUT</th><th>leak, calib-free</th><th>peak [mm]</th><th>over the gel</th><th>truncated</th><th>LUT vs calib-free shape</th></tr></thead><tbody>
+<tr><td>cnc_mini_26</td><td>120</td><td>0.016</td><td>0.013</td><td>2.71</td><td>14%</td><td>90%</td><td>+0.796</td></tr>
+<tr><td>cnc</td><td>120</td><td>0.015</td><td>0.014</td><td>1.71</td><td>1%</td><td>68%</td><td>+0.829</td></tr>
+<tr><td>feats</td><td>120</td><td>0.052</td><td>0.046</td><td>1.08</td><td>0%</td><td>99%</td><td>+0.688</td></tr>
+<tr><td>sparsh</td><td>90</td><td>0.038</td><td>0.031</td><td>0.53</td><td>0%</td><td>59%</td><td style='color:var(--bad)'>-0.082</td></tr>
+<tr><td>faf</td><td>119</td><td>0.044</td><td>0.035</td><td>2.14</td><td>10%</td><td>96%</td><td>+0.777</td></tr></tbody></table></div>
+
+<p class="dim">Leak is mean |depth| off-contact over peak — zero for a coherent
+surface. “Over the gel” counts peaks past the 4.25&nbsp;mm elastomer, possible
+only where the contact runs off the sensor and the depth is extrapolated.
+“Truncated” is a fact about the capture, not the method, and bounds what any
+reconstruction can know. The last column is the two reconstructions agreeing
+with each other, which is evidence neither invents the shape — not that either
+is right.</p>
+
+<p><b>Sparsh is the row that matters.</b> −0.082 means the two disagree
+entirely there, and the frames show why: both render a round sphere press
+wrongly, in orthogonal directions. The LED azimuths are per-sensor wiring, and
+these are ours — on Sparsh's own spheres a sphere comes out 3.3× elongated
+under our value and 1.5× under the one measured from its data. Its force ρ is
+0.93 either way, which is the whole reason this stage is scored separately.</p>
+
 <figure><img src="assets/panel_cnc_mini_26.png" alt="cnc_mini_26 — GelSight Mini, CNC presses, 0–20 N"><figcaption>cnc_mini_26 — GelSight Mini, CNC presses, 0–20 N · markerless · 8 samples. Columns 5 and 6 are the same surface from each reconstruction; the calibration-free one is drawn with relative height because its scale is not recovered.</figcaption></figure><figure><img src="assets/panel_fota_cnc.png" alt="FoTa cnc_Mini — GelSight Mini, CNC probes"><figcaption>FoTa cnc_Mini — GelSight Mini, CNC probes · markerless · 8 samples. Columns 5 and 6 are the same surface from each reconstruction; the calibration-free one is drawn with relative height because its scale is not recovered.</figcaption></figure><figure><img src="assets/panel_feats.png" alt="FEATS — marker gel, FEA force labels"><figcaption>FEATS — marker gel, FEA force labels · MARKER · 8 samples. Columns 5 and 6 are the same surface from each reconstruction; the calibration-free one is drawn with relative height because its scale is not recovered.</figcaption></figure><figure><img src="assets/panel_sparsh.png" alt="Sparsh / Meta — 10 gel pads"><figcaption>Sparsh / Meta — 10 gel pads · markerless · 8 samples. Columns 5 and 6 are the same surface from each reconstruction; the calibration-free one is drawn with relative height because its scale is not recovered.</figcaption></figure><figure><img src="assets/panel_faf.png" alt="FeelAnyForce — 42 captures"><figcaption>FeelAnyForce — 42 captures · markerless · 8 samples. Columns 5 and 6 are the same surface from each reconstruction; the calibration-free one is drawn with relative height because its scale is not recovered.</figcaption></figure>
 
 <details><summary>Three claims made here and withdrawn (with their

results.html

@@ -95,7 +95,7 @@ maximum.</figcaption></figure>
 position, ρ&nbsp;0.812 against the LUT's 0.763, MAE 1.024 against
 1.113&nbsp;N. It leads on four of the five ground-truth sets; the loss is
 FEATS, the marker gel React does not use.</p>
-<p>The two agree at ρ&nbsp;=&nbsp;0.868 over 1,800 React
-frames, mean difference 0.82&nbsp;N. <b>The published dataset still
+<p>The two agree at ρ&nbsp;=&nbsp;0.877 over 1,800 React
+frames, mean difference 0.84&nbsp;N. <b>The published dataset still
 carries the LUT column</b>: switching it needs 36 episodes reprocessed.</p>
 </div></body></html>