ChangChrisLiu/E3D-Desktop-Disassembly

E3D: End-of-Life Desktop Disassembly Dataset

E3D = End-of-life Desktop Disassembly Dataset.

Status: work in progress (October 2026). Raw captures are being uploaded; annotations are released incrementally as they are verified. Structure, file names and the annotation format may still change before the first tagged release. Licence and citation will be added with the first release; until then the data is provided for inspection only.

What it is

66 end-of-life (EoL) business desktop PCs (Dell OptiPlex / Precision, HP Compaq / EliteDesk / ProDesk / Z, one Lenovo, one Apple G4, one custom build) were disassembled by hand, one step at a time (~2 800 steps in total). After every step four cameras took a picture of the chassis:

view	camera	image	notes
scanner	structured-light scanner, top-down	1600 x 1600 PNG, 10 exposures per step (P_0 … P_9)	the main view: sharpest, no operator occlusion
oak Camera_1	OAK-D Pro, near top-down	12 MP JPEG + aligned 1280 x 800 PNG + depth (NPY / PNG) + PLY
oak Camera_2	OAK-D Pro, oblique	same	operator sometimes in view
realsense	Intel RealSense, oblique	1280 x 720 colour + depth + PLY

Each desktop also has "component" shots (the removed parts laid out on the bench) from the OAK cameras, the RealSense and, for the scanner, a separate all-components scan.

The capture log of every desktop (one row per step: which part was acted on) is included; it is the backbone of the annotation: a step's row names the part that leaves, opens or is unplugged at that step, so frame k shows the state after step k (frame 1 = the untouched machine).

Annotations (incremental)

Annotations are made with a purpose-built tool (reverse-order walk from the empty chassis back to the untouched machine) and consist of, per view and per frame:

• layered instance masks: every part has an amodal shape (the whole part, including what is hidden) and a z-order; the visible mask is the shape minus the layers above it;
• per-step part states (installed / loosened / open / displaced / removed …) derived from the log and corrected by the annotator;
• relations between parts (parent, attached, fastens, socket host, cable) and disassembly precedence constraints;
• visibility labels (visible / partially occluded / occluded by a cable / occluded by an unlabelled object / not visible from this viewpoint / outside the frame);
• vision-language question/answer records derived from the above.

annotations/scanner/ holds the scanner-view release of 2026-10-29: the 57 desktops whose scanner frames are all confirmed and conflict-free (2 473 frames; D01, D48, D52, D56, D61–D65 follow). Per desktop D<NN>/:

file	content
labels/S<nn>.png	16-bit instance-id map of the confirmed visible masks of frame S<nn> (0 = background; ids in instances.json)
annotations.json	COCO: per annotation the visible segmentation, bbox, area plus amodal_segmentation / amodal_bbox / amodal_area (the whole part, hidden pixels included), occluders, clean, visibility and state; parts covered completely are crowd rows
instances.json	the parts of the desktop: key, class, attributes, relations (parent / attached / fastens / socket host / cable)
frames.json	per frame: the raw picture it was drawn on (source_folder + file_name under raw/scanner/, sha256), ROI, pose segment, step name, the actions of that step, review status; frames left out and why
sequence.json	the step sequence: for every step the parts acted on, the verb and the resulting part states

annotations/scanner/README.md (written by the exporter) describes every field; clean_parts.csv lists the instances whose masks are unoccluded on every frame; manifest.json carries the sha256 of every file. The pictures themselves are not repeated here: frames.json points at raw/scanner/<source_folder>/<file_name>. taxonomy.yaml is the part taxonomy (classes, attributes, states, verbs) used by all annotations.

Layout

README.md
taxonomy.yaml                      part taxonomy, states and verbs
raw/scanner/<batch>/<batch>/<NN>/RGB<n>1/P_0..P_9.png     scanner captures of desktop NN (n = capture number)
raw/scanner/<batch>/<batch>_allcompoentsrgb/...            scanner all-components scans
raw/oak/Desktop NN/{Disassemble,Components,Side view,Calibration}/Camera_1|2/NNN/
raw/realsense/Desktop NN/
annotations/scanner/README.md      the exporter's description of the annotation files
annotations/scanner/D<NN>/{labels/S<nn>.png, annotations.json, instances.json, frames.json, sequence.json}

raw/ keeps the folders exactly as the capture rigs wrote them. Capture numbers and step numbers are not always identical (duplicate shots, a few missed shots); the mapping is part of the annotation release.

Provenance

Captured May–June 2025 at Texas A&M University. The room videos recorded during capture are not included yet; edited clips showing the operator's actions will be added later.