File size: 2,104 Bytes
8e6cae5
 
 
 
 
 
481f2f2
8e6cae5
f39bdd3
 
8e6cae5
481f2f2
f39bdd3
8e6cae5
a5436de
 
 
 
f39bdd3
a5436de
8e6cae5
 
 
 
a5436de
 
f39bdd3
 
 
 
8e6cae5
 
 
f39bdd3
a5436de
f39bdd3
8e6cae5
 
 
 
a5436de
8e6cae5
a5436de
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
# Reconstruction status

SonoIntent v1.0.0 is a **frozen benchmark release**: the rendered images,
masks, splits and texts are treated as fixed artifacts. This file states
exactly what can and cannot be reconstructed from published material.

## Reproducible from the released subset

The HC18, PSFHS, BrEaST, CCAUI, CUBS, FASS and Segthy annotations, images,
masks and index are published here.

- split membership and counts for the released subset
  (8,437 / 498 / 7,934 / 0);
- `public_id` uniqueness and per-task dataset / setting / tier labels;
- every `asset_image` and `asset_mask` path, with matching image and mask
  dimensions and binary mask values;
- content-hash equality of every released image and mask against
  `image_sha256` / `mask_sha256` in the annotations and index;
- loading all 16,869 tasks and running mask-level evaluation without access to
  the original source data.

## Reproducible with local access (maintainers only)

- existence and content-hash equality of every rendered file against the
  frozen build (requires the internal identifier mapping, which is not
  published);
- re-running the HC18 outline-to-mask fill check, the PSFHS class-to-task mask
  split, the BrEaST alpha-to-mask conversion, the CUBS manual-boundary fill,
  the FASS structure-to-task split, and the Segthy slice/label extraction
  against the source records.

## NOT reproducible from published material

- pixels for the other 23 sources, or for the full 130,928-sample benchmark;
- the frame extraction rules for video/volume/container sources (frame rate,
  start/end, dedup, naming, orientation), including the Segthy slice selection;
- the CAMUS intermediate-frame propagation algorithm and its QC
  (51,681 propagated masks);
- the ACOUSLIC interpolation pipeline (6,620 records);
- the target-to-mask naming/dedup rules and the per-file mapping generator;
- the question/reasoning template engine and the Full V1 assembly driver.

A one-command rebuild of the full benchmark is therefore **not claimed**. The
released subset is directly usable for evaluation, and its files are frozen.