OpenOSV 0.5.1

OpenOSV 0.5.1: Premiere on a car mount, fixed at the cause

0.5.0 was measured with osvtool. Premiere Pro renders a different path: the classical flow solver and the per-column seam table. Run on the same two 8K car drives through that path, the stitch still wobbled, stepped and popped, and the horizon still leaned. Every one of those had a root cause, and every one is fixed in the code Premiere, DaVinci Resolve and VEGAS Pro share.

What's fixed

• Light poles beside the seam are straight, and the seam no longer jumps every eight frames. The seam table took a shift from every column that matched at all, so blank sky handed it random shifts of up to ±4° that bent anything running along the seam, and a change rule turned that noise into a jump at every bucket start. Each column now carries a confidence, the table is a robust smooth of the confident columns only, and it steps at a bucket edge only when both measurements are sure. A lamp pole next to the 8K seam goes from 2.85 px of wobble to 0.19 px; the per-frame change at bucket starts from 3.4° on 16 % of the columns to 0.04° on none.
• Horizon Lock is level, in every host. 0.5.0 still left the car drives leaning 28° (sunset) and 7-11° (night): the camera's stored attitude was read with its axes transposed, and the near-static airborne sample could not tell. The new reading levels both drives within about 4° of their lamp posts and keeps the sunset sun at +6 to +10° through a 106° turn. Full and Smooth now turn with the car instead of against it. This is also the VEGAS "image skewed" report (#5). Saved projects with Pan keyframes on mounted clips will see their horizon move: that is the fix.
• The parallax correction aligns a car body a metre from the lenses. The classical solver could not reach the 2.4° offset a roof rail shows along the seam and left it doubled whenever it was accepted. It now searches along the seam direction first and moves only where the match is unambiguous: the car body's overlap match goes from 0.84-0.87 to 0.93-0.97, at about 10 ms more per bucket on the CPU.
• The correction no longer switches on and off from one moment to the next. Whether a moment used its 2-D correction depended on how much of the whole seam band's flow checked out, and on a car mount most of that band is sky. The check now counts only pixels with detail in both lenses, and a moment near the bar fades its correction out and the seam table in.
• The car's hood lines up across the seam even where the 2-D correction is accepted. The flow cannot follow the hood's 2.5° offset, so the seam table takes over the stretches of the seam the flow could not measure, where the table is at least 60 % sure and found an offset the flow missed. The hood's overlap match goes from 0.95 to 0.97 (median of 15 moments) and the whole seam band to 0.993; the night drive, where the table is unsure of the hood, keeps the flow's alignment; the 6K sample is untouched.
• Non-square projects (HDV, DV, anamorphic) are no longer stretched in VEGAS and Resolve, and the playback proxy smooths over the same seconds as the clip, so its framing stays within 0.1° of the final render.
• A damaged hardware-decoded proxy picture is decoded again in software instead of being delivered. A clip without a recorded focal length still opens, and a proxy never borrows its parent's. A ranged osvtool render plays its own sound.

What's new

• Hide Mount (Source Settings, and the OpenOSV Source generator in Resolve and VEGAS: On, Off, Auto). On a car, helmet or suction mount, Off gives the two lenses their full overlap back where the mask used to cut the seam (the mount can show); Auto measures the clip once and keeps the mask only where the lenses disagree. On is the default and what every existing project keeps, bit for bit.
• The plug-in logs now carry the library's and FFmpeg's messages, name the clip on every per-bucket line, and say how much of a moment's flow was consistent and at what strength it applies.
• Diagnostics for a proxy frame that shows the wrong picture (docs/PREMIERE.md, Troubleshooting): decode and delivery fingerprints, OPENOSV_VERIFY_HW_DECODE=1, OPENOSV_IMPORTER_LRF_SOFTWARE=1, OPENOSV_VERIFY_DELIVERY=1, and a five-minute check that tells the decoder, the stitch, the copy into Premiere and Premiere's own cache apart.
• A car-mount checklist in the same Troubleshooting section: what to set, what the log says, what to send.
• osvtool: --hide-mount on|off|auto, seam --json with per-column confidence and both flow shares, --dump-bands writes raw and table-corrected bands, --attitude-convention ...-rig, and render says which flow solver it runs and whether that is the plug-ins' one.

Good to know

• The lens overlap of the 6K sample clip renders byte for byte as in 0.5.0 on the default settings. Car-mounted clips change on purpose.
• Where a near part of the car crosses the seam inside the mask's arc, the roof edge still steps about 2.7° with Hide Mount On (2.3° with Off): that crossing has no lens overlap, and what remains is the car's own disparity along the seam. The checklist says what to send if your mount looks worse than that.
• The attitude reading was settled on two level car mounts and one airborne clip. A clip with real pitch and roll (a tilted tripod, a slow pan, handheld) would pin the last couple of degrees down; if you have one, open an issue.

How to update

Close your editor, unzip the zip for it and double-click Install.cmd.

Downloads

• OpenOSV-0.5.1-premiere-windows-x64.zip: Premiere Pro
• OpenOSV-0.5.1-resolve-windows-x64.zip: DaVinci Resolve (preview)
• OpenOSV-0.5.1-vegas-windows-x64.zip: VEGAS Pro 17.0 (experimental)

ffmpeg-9.0.1-source.tar.gz and ffmpeg-9.0.1-vcpkg-port.zip are the source of the FFmpeg build the zips ship (LGPL-2.1).

SHA-256

308e574d5f59f94594206a418d8b31da396ee68a1f34905401622dc82b7b60ed  OpenOSV-0.5.1-premiere-windows-x64.zip
c9a52b232edfe3ccca1b377ecd02ee283438ee335bd11e99b1331a3da1cd7d49  OpenOSV-0.5.1-resolve-windows-x64.zip
989b0d0c09e720d1acee959744ef750bb4c3562a8393447a83116fdd7d7d4b9d  OpenOSV-0.5.1-vegas-windows-x64.zip