Spotting teleprompter reflections in broadcast footage with ReCAP

A presenter delivers a polished headline in a Sydney news studio. The lights are set, the autocue is rolling, the director calls the shot live. Then, after wrap, someone in quality control notices a faint streak of mirrored text drifting across the talent's cheek. The fragment is too small for viewers to read, but it is unmistakably the script bleeding back through the prompter glass. For broadcasters operating across Australia, this is a recurring headache that needs a fix, and it is precisely the kind of artefact ReCAP is being trained to find.

ReCAP ingests live or recorded video streams and runs a cascade of analysis passes: face detection, logo spotting, quality assessment, and now, a specific module tuned for reflective artefacts coming from teleprompter rigs. The system timecodes every event, indexes those events for retrieval, and exposes them to editors and compliance reviewers through standard production interfaces. What was once an invisible production blemish becomes a searchable record ready to be redacted, blurred, or simply not aired.

Why teleprompter reflections slip through quality control

Teleprompter reflections behave erratically. They appear only when the camera axis crosses the prompter glass at a particular angle, then vanish when the presenter turns their head a few degrees. The script text moves with the camera's aperture, smeared into a thin band that often looks like a lighting flicker to a human reviewer scanning hours of footage. They are rare enough that operators stop looking for them, yet common enough that any given bulletin can contain one.

Australian newsrooms feel this pressure more than most markets. The ABC's Ultimo centre produces rolling coverage across multiple time zones, SBS works through multilingual bulletins with longer scripts, and regional affiliates along the east coast run extended breakfast blocks where any small artefact has time to make it to air. The Australian Communications and Media Authority expects broadcasters to maintain rigorous accuracy, so a reflected headline from a script for a later segment can create editorial confusion if it escapes the building.

Add in the speed of live-to-air turnaround and the problem compounds. Operators in the gallery are trained to watch for the obvious mistakes: wrong slug, wrong mic, wrong camera. A subtle reversed word ghosting across a cheek does not register, even on a careful pass. That is why an automated detector has a real role to play here.

The detection method behind ReCAP's reflection module

ReCAP's reflection handling treats the telecine glass as an optical surface, not a graphic source. The first pass looks for the visual signature of a partially transmitting pane: a soft falloff that tracks the presenter's silhouette, plus a faint specular band that moves coherently over multiple frames. That pre-filter pulls out candidate regions at roughly real-time throughput on a single GPU node, returning a manageable list of windows per second rather than a dense set of detections.

Each candidate is then run through an optical character recognition pass trained on mirrored and rotated text. Standard OCR engines struggle with reversed characters, so the ReCAP team fine-tuned its text recogniser on synthetic reflections generated from real script pages. The match yields a confidence score and a coarse reading of the visible words, which is useful because it tells an operator what the prompter was showing at that moment.

The hardest part of the module is the discrimination step. Lower-thirds, channel bugs, and graphics-on-glass transitions all look superficially similar. ReCAP closes that gap by also checking motion vectors: a true reflection drifts with the prompter glass and is anchored in the studio environment, whereas a graphic is keyed onto the moving image and follows the foreground subject. Anything that fails this consistency check is suppressed.

Event record fields generated by the reflection module:

That record is what feeds the indexing layer described in the next section.

Logging and indexing events as searchable metadata

Once a reflection event is recorded, ReCAP hands it to its indexing layer, which behaves much like the project's face and logo indexes. Each event carries a unique identifier, a normalised timecode in the production's chosen frame rate, and a link back to the source media. Indexes are stored in a small-footprint format that can sit alongside a Media Asset Management system without forcing a wholesale migration.

Operators query the index through a REST endpoint, an NLE plugin, or a simple web dashboard. A search of the term "prompter reflection" returns every flagged segment in the past 24 hours, sortable by show, by studio, or by the presenter who triggered it. Producers can confirm the result in seconds and decide whether to re-record, apply a localised blur, or pass the clip through a separate compliance pass.

For stations handling large catalogues of footage, this kind of bulk processing is now standard practice. The same approach used for watermark batch processing sits comfortably alongside ReCAP's own reflection tagging: both apply tagging logic to long runs of footage without an operator having to scrub through every tape.

Studio practice: Australian broadcasters and live feeds

Walk into an ABC control room at five in the arvo during a federal election, and you will see exactly why this module matters. Producers are juggling five incoming feeds, a graphics package, and a presenter who is half a sentence ahead of the autocue. Anything that surfaces a stray reflection after wrap is a quiet cost to the team, who already take the long days as par for the course.

SBS studios in Sydney's inner west add another wrinkle. Multilingual bulletins run scripts in two or more languages per presenter, which means the prompter stack is doing real work and the reflection surface is in front of the lens for longer. A reflected line of Vietnamese subtitles during an English-language segment is the kind of artefact that no human reviewer will spot until a sharp-eyed viewer files a complaint weeks later.

Regional stations from Wagga to Cairns face the same issue with thinner crews. An automated log means a single overnight operator can rerun the day's recording against the reflection index at 3am and walk into the morning bulletin already knowing which clips need a quick second pass. That local workflow, run properly, saves productions from being caught out by the kinds of subtle gaffes that travel well online.

Wiring reflections into post-production pipelines

ReCAP exposes flagged events through a small set of integration points so that they fit into the editors' existing tools. The most common path is an XML or AAF sidecar that drops into an Avid, Adobe Premiere, or DaVinci Resolve bin as a stack of subclips with markers already in place. Editors open the timeline, jump to the markers, and decide on the spot whether to blur, re-shoot, or leave the material alone.

The wider principle behind this approach is covered in the project's write-up on using recap to automate creation of shot lists for post production editing, where detection metadata is converted directly into editor-ready tasks. Reflection events fit the same shape: a bounded region, a confidence score, a timecode range, and a recommended action.

Common integration points used by Australian post-production stacks:

Each of these means that the reflection record does not sit in a silo. It travels with the footage through ingest, edit, compliance, and archive without anyone having to re-key information.

Performance trade-offs and what's coming next

Running an extra module on top of face and logo detection is not free. On a live feed at 1080i50, the reflection pass consumes roughly a fifth of a modern GPU when set to detect conservatively. Stations that already run a tight compute budget sometimes defer the reflection pass until the recording has been written, treating it as a post-air check rather than a live overlay. That trade-off is reasonable today and will shift as the model is pruned in future releases.

False positives remain the main calibration challenge. A polished studio desk under strong key light produces specular streaks that look, to the untrained classifier, similar to a prompter reflection. The team is addressing this by layering in infrared and depth hints from additional sensors, which will be available on the next-generation capture rigs being tested in Hamburg. False negatives are rarer, but the team is also retraining the OCR stage on more typographic variants to catch curved and italicised prompts.

Reading accumulated signals and committing to a clear action is the muscle memory of operators working in timing-led roles, whether the call is knowing when to cash out at casino or how to handle a flagged reflection during a live bulletin. Trust the data, weight the evidence, and commit. The current release ships with a configurable policy file so each station can tune the threshold to its own risk tolerance rather than inheriting a one-size-fits-all setting.

By the next consortium milestone the plan is to fold reflection detection into the same real-time graphics-allowance path that ReCAP already uses for logo compliance. Once that happens, the system will not just log and index the artefact, it will quietly suppress it at the playout layer, leaving editors to focus on the segments that genuinely need human judgement rather than the kind of stray text that slips past everyone's eyes.

Build the index gradually on a single production week first. Run ReCAP across those five bulletins, sit down with the team that worked them, and adjust the policy file until the false positive rate feels right. From there, scale the same configuration across sister shows and let the indexed history build a long-term record of which studios, lenses, and presenter combinations tend to produce reflections, so the next round of rigging decisions in those rooms is informed by data the team already has on file.