A recording is the wire itself — containers, headers, CRCs — and that stays the master. But the world around it speaks other dialects, and two cheap bridges connect them.
Export, for looking and for datasets: per-frame DNG (the header's CFA order, bit depth and geometry mapped into DNG tags, so any clip opens in darktable or RawTherapee), and bare-mosaic .npy — (n, h, w) uint16 stacks, the convention ML raw datasets use — one decode_frame away, made official. Derived artifacts, never masters: the export carries a note of what it lost (the container's self-description).
Import, the more interesting direction: DNG → containers, so anyone's raw photos from any camera replay onto the wire and through an FPGA pipeline — sensor data from cameras nobody on the project owns. Bare-mosaic import too, for datasets: array in, CFA order and bit depth declared once at the boundary, self-describing containers out.
Done means: round-trips proven — container → DNG → container and container → mosaic → container bit-exact on the payload; an imported third-party DNG replayed onto the physical wire and verified through check; refusals with names for DNGs whose CFA or bit depth the registry does not carry.
What it unlocks downstream: every existing raw dataset and every camera that writes DNG becomes a source for receiver and pipeline development — the no-sensor story, extended to sensors nobody has.
This is a sponsorable capability target (see the funding section). Sponsored work lands in the open tree immediately, licensed like everything else — sponsorship buys ordering and named credit, not exclusivity. Scope by email: s.rabykin@gmail.com.
A recording is the wire itself — containers, headers, CRCs — and that stays the master. But the world around it speaks other dialects, and two cheap bridges connect them.
Export, for looking and for datasets: per-frame DNG (the header's CFA order, bit depth and geometry mapped into DNG tags, so any clip opens in darktable or RawTherapee), and bare-mosaic
.npy—(n, h, w) uint16stacks, the convention ML raw datasets use — onedecode_frameaway, made official. Derived artifacts, never masters: the export carries a note of what it lost (the container's self-description).Import, the more interesting direction: DNG → containers, so anyone's raw photos from any camera replay onto the wire and through an FPGA pipeline — sensor data from cameras nobody on the project owns. Bare-mosaic import too, for datasets: array in, CFA order and bit depth declared once at the boundary, self-describing containers out.
Done means: round-trips proven — container → DNG → container and container → mosaic → container bit-exact on the payload; an imported third-party DNG replayed onto the physical wire and verified through
check; refusals with names for DNGs whose CFA or bit depth the registry does not carry.What it unlocks downstream: every existing raw dataset and every camera that writes DNG becomes a source for receiver and pipeline development — the no-sensor story, extended to sensors nobody has.
This is a sponsorable capability target (see the funding section). Sponsored work lands in the open tree immediately, licensed like everything else — sponsorship buys ordering and named credit, not exclusivity. Scope by email: s.rabykin@gmail.com.