No-contact reference frames — how they're chosen & used
A GelSight difference image / contact metric is computed relative to a no-contact reference (the resting gel). This folder ships one per session and documents how to pick a per-episode reference, how to build the no-contact pool, and the caveat for a reference latent.
How the reference is selected
Each episode is a cut from one continuous per-day recording (episodes.jsonl
date), so all episodes of a day share the same gel & lighting.
- Primary — per-episode fuzzy mode (
toolbox.reference.fuzzy_mode_background): the per-pixel temporal mode of the episode is the resting gel even when a pixel is in contact part of the time (robust past the median's 50%-occupancy limit). Drift-free (it is the episode's own gel state). Preferred whenever the episode has clean frames — 63/64 sides. - Fallback — session reference (
toolbox.reference.session_reference, shipped as<date>/tactile_{left,right}.png): for a clip that is in contact its entire length its own mode is contaminated. Pool the whole day (any pixel is untouched most of the day) to recover the rest gel, then use it to label that episode's own clean frames. Example:2026-05-19 ep003 left(121 frames, all contact vs its own mode) → the session reference finds 26 real no-contact frames in it (indices 17–21, 53–73).
Do not use
get_reference(mode="p01")(quietest single frame): it can pick a marginal or contact frame, and returns a contact frame outright when no clean frame exists — inverting the difference image. Audit: 2/64 p01 anchors were contact frames.
Caveat: the session reference carries intra-day gel drift, so it differs from
a given episode by a diffuse global amount (~0.37 area, no object — see
drift_check.png). Use it only to label all-contact episodes; prefer the
per-episode mode otherwise.
How to use it
Shipped references (ready to use, native 640×480):
per_episode/<date>/<episode>_<side>.png— the per-episode rest gel (primary; fuzzy mode of that episode). One per episode-side.ep003 left(all-contact) ships its session reference instead (seeper_episode/manifest.json).<date>/tactile_{left,right}.png— the per-session rest gel (fallback / day-level).
import imageio.v2 as imageio
from toolbox import load_video, difference, contact_metrics, l2_diff
import numpy as np
frames = load_video("data/motherboard/videos/2026-05-11/episode_016/tactile_right.mp4")
ref = imageio.imread("data/motherboard/reference/per_episode/2026-05-11/episode_016_right.png")
diff = difference(frames[t], ref) # Sparsh-style signed difference image
m = contact_metrics(frames[t], ref) # {intensity, area, mixed}
area = np.array([(l2_diff(f, ref) > 8).mean() for f in frames])
pool = np.where(area < 0.01)[0] # the episode's no-contact reference pool
To recompute a reference yourself (e.g. at another resolution / with in-bin
refinement) use fuzzy_mode_background(frames) (per-episode) or
session_reference([...]) (per-day); toolbox/build_session_references.py
rebuilds the whole set. Shipped PNGs are the per-pixel mode (bin width 8).
from toolbox import load_video, fuzzy_mode_background, session_reference, difference, contact_metrics, l2_diff
frames = load_video(paths["tactile_left"]) # (T,H,W,3) uint8
ref = fuzzy_mode_background(frames) # (H,W,3) rest gel
diff = difference(frames[t], ref) # Sparsh-style signed diff
m = contact_metrics(frames[t], ref) # {intensity, area, mixed}
# no-contact pool = frames whose gel is at rest:
area = np.array([ (l2_diff(f, ref) > 8).mean() for f in frames ])
pool = np.where(area < 0.01)[0] # reference-frame pool
Reference latent — the whole window must be no-contact
If you encode with a video tokenizer (e.g. Cosmos), a frame's latent depends
on its entire temporal window (Cosmos CV4x8x8: ti=0 is a single frame, ti≥1
encodes a 4-frame group). A frame can itself be no-contact while its window/group
straddles a contact boundary — then the latent bakes in contact even though
the frame is clean. Audit: this inflated one episode's background floor from 0.05
to 0.11 and produced a spurious contact "artifact" in every delta map.
Rule: for a clean reference latent, require every frame in the encoder's window to be no-contact, not just the anchor frame:
# a latent frame (window start s, group g) is a valid reference iff the WHOLE
# group is no-contact:
group = [s] if ti == 0 else range(s + 1 + (ti-1)*4, s + 1 + ti*4) # Cosmos ti map
valid = all(area[r] < 0.01 for r in group)
For an image tokenizer this is unnecessary (each frame encodes independently); a
Cosmos ti=0 latent is equivalent to a single-image encode (verified to 0.8%).
Preview — 20 reference frames per episode
previews/<date>/<episode>_<side>.png shows 20 frames sampled from each
episode-side's no-contact pool (labelled f<frame> a<area>). Every episode-side
has a real, non-empty pool (median ~2145 frames, ~56% of the clip; min 26 for
ep003 left via the session fallback). Pool sizes are in previews/pool_manifest.json.
Preview gallery — 20 reference frames per episode-side
Each grid = 20 frames sampled from that episode-side's no-contact pool (f<frame> a<area>). pool= is the full pool size; via = which reference labeled it.
2026-05-10
episode_000 left — pool=3534 (episode mode)
episode_000 right — pool=4762 (episode mode)
episode_001 left — pool=4405 (episode mode)
episode_001 right — pool=5799 (episode mode)
episode_002 left — pool=6606 (episode mode)
episode_002 right — pool=7455 (episode mode)
episode_003 left — pool=4653 (episode mode)
episode_003 right — pool=4979 (episode mode)
episode_004 left — pool=908 (episode mode)
episode_004 right — pool=591 (episode mode)
episode_005 left — pool=342 (episode mode)
episode_005 right — pool=477 (episode mode)
episode_006 left — pool=1705 (episode mode)
episode_006 right — pool=1133 (episode mode)
episode_007 left — pool=1092 (episode mode)
episode_007 right — pool=1039 (episode mode)
episode_008 left — pool=769 (episode mode)
episode_008 right — pool=760 (episode mode)
episode_009 left — pool=1154 (episode mode)
episode_009 right — pool=807 (episode mode)
episode_010 left — pool=403 (episode mode)
episode_010 right — pool=583 (episode mode)
episode_011 left — pool=915 (episode mode)
episode_011 right — pool=743 (episode mode)
2026-05-11
episode_003 left — pool=6550 (episode mode)
episode_003 right — pool=7148 (episode mode)
episode_004 left — pool=1786 (episode mode)
episode_004 right — pool=1844 (episode mode)
episode_005 left — pool=8123 (episode mode)
episode_005 right — pool=9523 (episode mode)
episode_006 left — pool=7565 (episode mode)
episode_006 right — pool=7654 (episode mode)
episode_007 left — pool=4624 (episode mode)
episode_007 right — pool=3903 (episode mode)
episode_008 left — pool=8769 (episode mode)
episode_008 right — pool=5793 (episode mode)
episode_009 left — pool=2718 (episode mode)
episode_009 right — pool=1850 (episode mode)
episode_010 left — pool=4491 (episode mode)
episode_010 right — pool=4117 (episode mode)
episode_011 left — pool=5060 (episode mode)
episode_011 right — pool=4518 (episode mode)
episode_012 left — pool=9042 (episode mode)
episode_012 right — pool=7426 (episode mode)
episode_013 left — pool=9224 (episode mode)
episode_013 right — pool=7176 (episode mode)
episode_014 left — pool=1413 (episode mode)
episode_014 right — pool=1044 (episode mode)
episode_015 left — pool=3881 (episode mode)
episode_015 right — pool=3433 (episode mode)
episode_016 left — pool=6756 (episode mode)
episode_016 right — pool=7268 (episode mode)
episode_017 left — pool=636 (episode mode)
episode_017 right — pool=2441 (episode mode)
2026-05-19
episode_000 left — pool=289 (episode mode)
episode_000 right — pool=358 (episode mode)
episode_001 left — pool=149 (episode mode)
episode_001 right — pool=295 (episode mode)
episode_002 left — pool=568 (episode mode)
episode_002 right — pool=575 (episode mode)
episode_003 left — pool=26 (session ref (fallback))
episode_003 right — pool=65 (episode mode)
episode_004 left — pool=52 (episode mode)
episode_004 right — pool=90 (episode mode)