---
license: cc-by-4.0
task_categories:
  - image-classification
  - feature-extraction
  - video-classification
tags:
  - tactile
  - gelsight
  - gelsight-mini
  - video
  - sequence
  - temporal
  - tactile-video
  - pretraining
size_categories:
  - 100K<n<1M
pretty_name: GelSight Mini Pretrain · Video
---

# GelSight Mini Pretrain · Video / Sequence Subset

> 🎬  **Companion to [`yxma/gelsight-mini-pretrain`](https://huggingface.co/datasets/yxma/gelsight-mini-pretrain).**
> Where the main repo treats every kept frame as an independent image, this repo
> **preserves temporal sequences** — one row per frame, ordered, with explicit
> sequence-id + position metadata, for **video tactile pretraining**.

## Why this repo

The main repo's pipeline applies **perceptual-hash dedupe** within each capture
to drop near-identical adjacent frames. That's great for image-level
pretraining (no redundancy) but **bad for video-level pretraining** (the small
inter-frame changes are exactly the temporal signal we want to learn).

This repo:
- Uses the **same area+intensity validity filter** as the main repo
  (drops gel-at-rest frames — `A ≥ 40, I ≥ 10, PIXEL_THRESH = 10`)
- **Skips phash dedupe** so consecutive frames within a sequence are preserved
- Adds **sequence-level metadata** so users can reconstruct clips

## Schema additions (vs main repo)

| Column | Type | Description |
|---|---|---|
| `sequence_id` | string | unique identifier for the sequence the frame belongs to |
| `frame_in_seq` | int32 | 0-indexed position of this frame within the sequence |
| `sequence_length` | int32 | total active-contact frames in this sequence (helpful for windowing) |
| `fps` | float32 | original capture frame rate (typically ~25 for GelSight Mini) |

Existing columns from the main schema (image, capture, split, height, width,
source, markered, domain, etc.) are all carried through.

## Sources (video sources only)

| Source | Sequence unit | # sequences | Active frames |
|---|---|---:|---:|
| `gelslam` | one `gelsight.avi` per episode (tracking + reconstruction) | 155 | TBD |
| `tactile_tracking` | one `gelsight.avi` per trial | 84 | TBD |
| `fota_unlabeled` | one capture (object × pose × side) | 60 | TBD |
| `real_tactile_mnist` | one touch-window per touch | 153,600 | TBD |

(Counts will be filled in as each source is processed.)

## Recommended uses

- Video/temporal tactile pretraining (X-frame clips)
- Slow-motion contact dynamics modelling
- Sim-to-real transfer learning over time
- Cross-frame contrastive learning (within sequence vs across sequence)

## Sample-windowing recipe

```python
from datasets import load_dataset
import numpy as np

ds = load_dataset("yxma/gelsight-mini-pretrain-video", "gelslam", split="train")
# Group rows by sequence_id, sorted by frame_in_seq
import pandas as pd
df = ds.remove_columns(["image"]).to_pandas()  # just metadata
seqs = df.groupby("sequence_id")["frame_in_seq"].count()

# Pull a 16-frame clip starting at the first frame of episode "0042"
clip = ds.filter(lambda r: r["sequence_id"] == "0042" and r["frame_in_seq"] < 16)
frames = [r["image"] for r in clip]  # PIL.Image list of length 16
```

## Status

- Schema established
- Sources processing in progress; counts TBD

## License

CC-BY-4.0 (same as main repo). Cite upstream sources per `SOURCES.md` in the
[main repo](https://huggingface.co/datasets/yxma/gelsight-mini-pretrain/blob/main/SOURCES.md).