Repository navigation
Releases
CVPR-Figure 3.1.0
Auditing all 17 figures of a real CVPR submission with 3.0 turned up three
defects in the auditors themselves: one problem they missed, one false
alarm, and one bad side effect of 3.0's own subscript fix. Each was measured
in LibreOffice and Keynote before it was changed.
Fixed
- Shadows inherited from the theme went unreported. FigGenie's PPTX gives
every shape an empty<a:effectLst/>plus<p:style><a:effectRef idx="1|2">,
which points at theme effect styles that draw a 35–38 %outerShdw. By the
OOXML spec the empty list cancels the shadow, and PowerPoint and Keynote
draw none. LibreOffice ignores the override and draws all 139 shadows, then
exports each one as a blurred 97–313 ppi grey picture.audit_pptx.pytook
the override at face value. It now resolveseffectRefagainst the theme
the slide master actually uses, and reportsraster-effect (theme)as an
error, whatever the shape writes. - Class legends for rendered images were reported as orphaned. A legend
for occupancy or segmentation classes (car, bus, vegetation …) explains
colours that appear only inside embedded renders. Every qualitative figure
of the submission got 10legend-orphanerrors. Both auditors now look
inside the pictures:audit_pptx.pychecks the embedded media,
validate.pythe files in image slots. A pixel counts for a legend colour
when it is that colour times a shading factor between 0.45 and 1.2, so
shaded voxels are matched. Calibration on Fig. 4's 24 renders: every legend
class covers at least 0.2 % of some render, while clearly absent colours
stay at or below 0.006 %. The threshold is 0.1 %. Grey and pale colours
are never looked for, because backgrounds would match them. In a legend
whose other entries are found, a class that occurs in no picture is
reported once, as a warning. Keeping a full palette across figures is a
legitimate choice. Needs Pillow or PyMuPDF; without either, the auditors
say so rather than guess. - Enlarged subscripts dropped too far in LibreOffice. LibreOffice prints a
baseline-shifted run at 58.5 % of its size and moves it by baseline × the
run's own size. So 3.0's 15 % larger scripts at −25 % sat 0.29 × body
below the line. Scripts enlarged by hand to 1.25× sat 0.31 × body below,
43 % of their own height, and a loned₁looked likedand1. The
writer now sets the shift in body terms and divides it by the enlargement:
−17 % for subscripts (0.20 × body, the measured best being 7.8 pt on
7.0 pt at −18 %), +28 % for superscripts (0.32 × body, as in the SVG).
Added
audit_pptx.py --fix OUT.pptxwrites a repaired copy, then audits it.
Theme effect references are set to 0 and the theme's effect styles are
emptied. Subscripts LibreOffice would print detached are re-shifted.
Every other byte is copied unchanged. On the submission's Fig. 2 it raised
27 subscripts from −25 % to −16 %; on FigGenie's original it removed all
139 shadows, and the LibreOffice PDF went from 157 images to 18.script-too-low: inaudit_pptx.py, a subscript shifted more than about
−20 %, which LibreOffice prints over 35 % of its own height below the line.
Inaudit_pdf.py, the same drop measured in the PDF.audit_pdf.pyraster-shadow: images that are one flat colour under a soft
mask are reported together as rasterised shadows or transparency, instead of
as oneimage-low-ppiper shadow plus apage-is-raster.audit_pptx.py --legend-in-imagesand the legend item keyin_image: true
declare that a colour is drawn inside the pictures.audit_pdf.py --source fig.pptx, passed automatically byexport_pptx.py:
a picture that was already JPEG in the PPTX is a note, not animage-jpeg
warning. So is a JPEG at the native frame size of a common driving camera
(nuScenes 1600×900, Waymo, KITTI, Argoverse 2, Cityscapes, DDAD, ONCE …).
Camera photos are JPEG at the source.audit_pptx.pyreads the theme through the slide → layout → master chain,
not the firsttheme*.xmlin the package.
CVPR-Figure 3.0.0
Everything in this release comes from using 2.0 on a real CVPR submission.
The figure left the engine clean; the defects appeared afterwards, in the PPTX
a person edited in WPS and in the PDF exported from it. So the audit now
follows the files that ship, every symbol must be defined, type is larger, and
the PPTX writer was rebuilt to survive hand editing.
Added — auditing what ships
scripts/audit_pptx.py— audits any PPTX (this engine's, FigGenie's, or a
hand-drawn one) at its printed width:short-final-segment(bent arrows WPS
draws at 45°),unglued-arrow,freeform-arrow,glue-broken,
text-too-small,lost-subscript(against--specor a source--ref
SVG, or heuristically:z1:4,d1),legend-orphan,image-low-ppi,
raster-effect,no-background,off-slide,mixed-fonts, and the
arrowhead size each application will draw. On the hand-edited overview of
a real submission it reports exactly the 2 short final runs, 20 flattened
subscripts and 1 orphaned legend entry that had reached the paper.scripts/audit_pdf.py— the PDF LaTeX includes: glyph sizes at the printed
width, image ppi and JPEG/JPX encoding, unembedded and Type 3 fonts, pages
that are one raster. Needs PyMuPDF.scripts/export_pptx.py— lossless PPTX → PDF through LibreOffice (no
downsampling, lossless compression, quality 100) or Keynote (Best), followed
byaudit_pdf.cvprfig/ooxml.py— reads a PPTX back: absolute line paths through group
transforms, flips and rotation; arrowheads; text runs with sizes and
baselines; pictures with pixel and printed size; glue; effects.- Rule, in
SKILL.md,stance.mdandafter-handover.md: once the PPTX is
edited it is the only source — export from it, never re-render over it.
Added — symbols and captions
$...$maths in every label: italic variables, upright digits and
operators, TeX spacing around binary operators and relations, accents,
\mathrm/\mathbf/\mathbb, Greek, primes,\log-style names;\$for
a literal dollar.symbols:table andcaption:/caption_tex: file.tex#label; checks
undefined-symbol,symbol-not-in-caption,symbol-label-bare,
symbol-defined-elsewhere,upright-variable,symbol-style,
unused-symbol,caption-long. Operators (⊕ ⊗ ⓒ), stacking brackets and
repeat badges (×L) count as symbols.paper:— onepaper.yamlper paper holding venue, font, role colours and
the symbol table.contract.mditem 7, the symbol table; caption template gains a fourth
sentence (define every symbol, in reading order) and a note to recompile and
check the page count.from_code.pydrafts declare the operators they draw and start a caption
that defines them.
Added — routing and the PPTX writer
- Glued connectors:
straightConnector1/bentConnector2/
bentConnector3withstCxn/endCxn, solved for rotation and flips; custom
polygons carry connection sites where the arrows meet their outline.
LibreOffice re-opens every template's PPTX with all 159 arrows where the
engine put them (tests/compat/lo_roundtrip.py). Routes with three or more
bends stay freeform and are listed. - The run after the last bend is at least 2.5 arrowheads; when the geometry
cannot allow it the auditor says so and the PPTX gives the head its own
straight segment. - Lanes: dog-legs that would jog on top of each other are spread 4 pt apart;
edges sharing a port share one trunk; stage gaps widen until their wires fit
(corridor-too-narrow,edges-share-corridor). - Arrowhead presets chosen for the application that exports the PDF
(pptx_app: office | wps | keynote) and drawn identically in the SVG;
pptx_heads: shapefor identical triangles everywhere;
pptx_connectors: freeformfor Keynote. - The legend is emitted to PPTX and VSDX at all (it was missing), as one group
with a sub-group per entry; legend items honourdash:andarrow:. - Explicit white slide background (Keynote rendered our slides black).
- Script runs written 15 % larger to offset the applications' own reduction;
whole-point type forpptx_app: keynote; slack on text boxes. type_scale,align_x/align_y,label_rotate,fixed_gaps.
Added — auditor
edge-crosses-text, diagonal-edge, label-overlap, text-off-canvas,
edge-off-canvas, legend-unused-entry, unexplained-line-style,
legend-overlap, ai-palette (Tailwind / Material / Bootstrap stock
colours), saturated-fill, tinted-outlines, arrow-rainbow,
image-low-ppi / image-soft / missing-image, module-label-small,
script-too-small; legibility now covers every text element, not only module
labels.
Changed — recalibrations
- Type. Body 7.0 → 7.5 pt, stage titles 7.8 → 8.5, captions 6.5 → 7.0,
annotations 6.0 → 6.5. Floor 5.0 → 6.0 pt, warning 5.6 → 6.5 pt, module names
≥ 7.0 pt — measured on the template papers users hold up as the model
(SparseWorld 7.5 pt median, Occ3D 8.2, BEVDepth 8.6); see
corpus-report.md. - Roles.
core/oursgold → orange,decoderpurple → green,encoder
green → blue,outputblue → green,temporalorange → gold,auxteal
light → pale: blue in, green out, orange ours, grey inherited. - Strokes and heads.
flow0.45 → 0.60 pt; arrowheads 3.97 × 3.97 pt
(LibreOffice/PowerPointsm/sm),port_stub5 → 9.9 pt. - SVG→PDF now prefers LibreOffice with lossless options; rsvg-convert
mis-places the glyph after an italic "f" in Times New Roman on macOS.
Fixed
- A caption wider than its shape stretched the shape and moved its ports to
the caption's edge, so arrows started in empty space. Shapes keep their
size; ports sit on the shape; a captioned row centres the shapes; the south
port sits below the caption. - Ports on slanted shapes (trapezoid, parallelogram, plane) now lie on the
outline. - A stage narrower than its title let the title spill off the canvas.
- Nested containers with a title above it did not reserve room for it.
- Arrows into a titled stage ran through the title.
anchor: startlabels were drawn from the box centre; notes ignored
anchorin PPTX and VSDX.- Leading blanks of a run were dropped by converters, collapsing math spacing.
- The VSDX writer printed inline markup literally (
*α*,**0.00**) and
emitted no legend. - Edge labels in PPTX/VSDX were placed differently from the SVG.
e → eroutes ran through the target box.- Facing ports a few points out of line produced 2 pt jogs; they are now one
straight arrow.
Templates
All sixteen templates and the quickstart were reworked for the new type and
roles, carry a symbol table and a claim-style caption, and audit with 0 errors
(the remaining warning is the empty image slots). Two had real defects the new
checks found: teaser-comparison listed two legend entries for lines it does
not draw; gated-module aligned gates with hand-measured spacers that broke
when the type grew (now align_x).