license: cc-by-nc-4.0
task_categories:
- video-classification
- visual-question-answering
language:
- en
tags:
- laboratory
- life-science
- protocol-compliance
- egocentric-video
- biology
- wet-lab
size_categories:
- n<1K
configs:
- config_name: XMglass
data_files:
- path: XMglass/xm.csv
split: train
- config_name: DJI
data_files:
- path: DJI/dji.csv
split: train
LSV: LabSuperVision Benchmark
Dataset Description
LSV is a multi-view video dataset of wet-lab biology experiments, captured from both first-person (XMglass smart glasses) and third-person (DJI action camera) perspectives. Each video records a researcher performing a laboratory protocol and is annotated with the corresponding protocol text, scene type, and—where applicable—deliberate procedural errors.
The dataset is designed for research on:
- Protocol compliance monitoring — detecting whether a procedure was followed correctly
- Procedural error detection — identifying specific deviations from standard protocols
- Egocentric video understanding — understanding lab activities from a first-person view
- Video-language grounding — linking protocol text to video segments
Dataset Structure
LSV/
├── XMglass/
│ ├── xm.csv # Metadata (90 entries)
│ ├── XMprotocol/ # Protocol text files (22 files)
│ └── XMvideo/ # Video files (105 files, ~75 GB)
├── DJI/
│ ├── dji.csv # Metadata (161 entries)
│ ├── DJI-Protocol/ # Protocol text files (17 files)
│ └── DJI-Video/ # Video & image files (251 files, ~219 GB)
Metadata Fields
Both CSV files share the following columns:
| Column | Description |
|---|---|
Slice_ID |
Unique identifier (e.g., XM_001, DJI-001) |
Exp_ID |
Experiment group identifier |
Date |
Recording date |
Video Name |
Filename of the video/image |
Scene |
Recording location (TC hood, bench, TC room, TC) |
Operation |
Description of the procedure performed |
Protocol |
Filename of the corresponding protocol in the protocol folder |
Issue (if any) |
Description of intentional procedural errors, if present |
Length |
Duration of the video |
Time_stamp |
Timestamps of protocol steps within the video |
Tools |
Lab equipment used |
Data Collection
XMglass (First-Person View)
- Device: XM smart glasses with built-in camera
- Entries: 90 annotated video clips
- Scenes: Tissue culture (TC) hood, bench, TC room
DJI (Third-Person View)
- Device: DJI action camera
- Entries: 161 (127 videos + 34 images)
- Scenes: TC hood, bench, TC room
- Note: Some experiments include paired first-person and third-person recordings of the same procedure
Covered Procedures
The dataset covers a range of common molecular biology and cell culture techniques, including:
- Cell line passaging and seeding (HEK293T, iPSCs, cancer cell lines)
- Lentiviral packaging, collection, and infection
- CRISPR/Cas9 delivery
- PCR reaction setup and colony PCR
- Serial dilution
- DNA gel electrophoresis (E-gel loading)
- RNA extraction
- Cell freezing and thawing
- Restriction digestion, Gibson assembly, Golden Gate reaction
- Transformation
- MiniPrep and NanoDrop quantification
- FACS staining
Error Annotations
Many videos include deliberate procedural errors with detailed descriptions. Examples:
- Skipping a pipetting step
- Not changing pipette tips between reagents
- Adding reagents in the wrong order
- Omitting incubation or mixing steps
- Forgetting to add a critical reagent
These error annotations enable benchmarking of automated protocol-compliance systems.
Usage
from datasets import load_dataset
# Load XMglass metadata
xm = load_dataset("YinkaiW/LSV", name="XMglass", split="train")
# Load DJI metadata
dji = load_dataset("YinkaiW/LSV", name="DJI", split="train")
License
This dataset is released under the CC BY-NC 4.0 license.