Before the kit changes hands
A repair kit comes back. The case looks full. Is the adapter present, hidden beneath a cable, or different from the one that went out?
Countback gives that inspection a visual workspace. Built for repair benches and equipment-loan desks, it helps an operator compare known items across photographs, investigate uncertainty, and leave a review the next person can follow.
Watch the working prototype · Download the application and judge package · Technical report
What you can do today
- Set the reference. Name the items to inspect, add close-up reference photographs, and supply up to three ordered photographs of the same returned arrangement.
- Investigate across views. OpenCV 5 analyzes visual evidence. When an item remains unresolved, the controller can inspect the next supplied photograph and record the decision.
- Take a closer look. The Review Desk presents the reference, a tentative inspection focus when available, and the full scene.
- Make the handoff. Record an assessment and reason, save and reopen the review, and export a readable inspection handoff.
Human assessments begin pending and remain separate from machine evidence. The present release analyzes supplied photographs; final acceptance stays with the operator.
How OpenCV changes what happens next
The loop is observe → assess → inspect another supplied view when needed → review. An OpenCV result determines whether the controller analyzes the next supplied photograph, and the trace explains that decision.
The optional viewpoint-aware path estimates foreground, simulates affine viewpoints, and matches distinct SIFT landmarks. Eligible matches become tentative focuses. When the evidence is insufficient, the operator sees the full scene.
The important distinction is between an item that was not found and an item that is actually absent. Countback preserves that uncertainty. A matching patch also cannot establish unique identity, physical quantity, or kit completeness.
What the prototype has demonstrated
Photographic evaluation. A preselected cohort used 18 RGB photographs from nine clips in two environments, with the same nine enrolled objects. The conservative affine-only path produced 5 localized focuses on 24 visible-target queries and no focus on 30 absent-target queries. Earlier paths produced more visible-target regions but also many distracting highlights when the target was absent.
That trade-off led to the optional focus-or-full-scene display. Useful coverage remains low, particularly for low-texture objects. The cohort contains related frames and shared objects, and informed the display policy; it is not an independent identity-accuracy or operator-benefit study.
Protocol, baselines, and complete results
Actual AWS execution. The original OpenCV 5 handler ran in a private, digest-pinned AWS Lambda container: four valid image analyses and five expected input refusals passed. One two-view public-photo request took 5.265 seconds end to end; its results remained appearance-only. The local viewpoint refinement was separate from that trial.
The extracted application package was checked across Python, JavaScript, browser interaction, public-photo integration, and packaging. The release includes the application, report, architecture, evaluation records, and checksums.
Next: a camera-guided return bench
Planned extension; not included in the current application or video.
The next version should help collect the evidence that is still missing. A printed reference mat and OpenCV marker detection would align checkout and return views, so the operator can inspect the same regions despite camera movement.
The target demonstration: remove one tool, cover another, and change a third. Countback would highlight regions for inspection, request a fresh photo after an item is uncovered or repositioned, and show how that action changes the evidence. The resulting review would connect the original condition, the follow-up observation, and the operator's decision.
This is the next engineering direction: vision that guides a useful physical action and then checks its result. Evaluation should measure resolved inspection tasks, incorrect clearances, and review time on held-out kits, including glare, shadows, occlusion, and similar-looking objects. An occupied slot alone would not prove the correct item is present.
Try it
Download the judge package and extract the nested Countback-Local-Candidate.zip intact. With Python 3.13, create and activate a virtual environment in countback-workbench, install requirements-workbench.txt, and run python pose_workbench.py. Open the printed local address. Use python workbench.py for the original mode.
The local workflow needs no AWS account and requires explicit processing permission. Private development photographs are excluded from the release and AWS trial. The temporary AWS deployment was removed after verification. A live judge screen-share was requested and is not yet confirmed; the runnable package and recorded demo are available now.
Original code: Joseph Ayanda, MIT, with AI assistance. Public images: UW-IS Occluded Dataset v1, CC BY 4.0, by Ekta U. Samani, Xingjian Yang, Srivatsa Grama Satyanarayana, and Ashis G. Banerjee. The 3:37 video uses actual local footage, neural narration, one scripted review note, and a labelled summary of AWS receipts.
Log in or sign up for Devpost to join the conversation.