I have read this post that answers with 4 commercial players that implement frame-accurate seeking and I will investigate them, but we are trying to build a tool for labeling objects in a video and we need to know precisely the frame that they have labeled so a subsequent watcher can validate their label.

We ask the users to watch until they see our desired object. Then pause the video and navigate to the "Best Frame". We give them buttons to go back and forward (1 second/Average FPS) to go move one frame at a time (unless our video has dropped frames, then they must push those buttons once for each dropped frame.)

[Bonus question: can we tell if gave us a new frame or just the same frame with an updated time-in-video? If we got the same frame, we could loop until the first new frame so the user wouldn't have to keep pressing the key!]

When they arrive at the best frame, we can record the time of that frame, but we are finding that the time often comes back with different milliseconds each time someone tags the exact same frame. So we can't reliably have a subsequent user seek the exact same frame to validate the first watcher's label. We are using Canvas to allow the first user to draw a bounding box around our object of interest. Yet we can't reliably redraw the bounding box if it's on the wrong frame as our objects move quickly.

Ideally, if it can't be a sequential frame number, can it be PTS code? Is there any other way we can determine the precise frame each time? We are building this for a controlled audience that we can train, so creative solutions are acceptable.

We are open to a "cheat" because we control each video in our workflow. If we can write "something" into the file that can be read by the HTML-Video, we could seek to the approximate time minus 0.5 seconds and then read the subsequent frames for "something" that can tell us we have an exact match.