Submit per-frame camera intrinsics predictions for evaluation on the withheld test sets of InFlux and InFlux++ Real. Submissions are prepared and uploaded with the InFlux command-line tools.
| Target | version |
Evaluated data |
|---|---|---|
| InFlux | influx |
Original InFlux test set |
| InFlux++ Real | influx_pp_real |
InFlux++ Real test set |
| Combined | all |
Both test sets, with separate and aggregate results |
For each camera parameter q ∈ {fx, fy, cx, cy} , we compute relative error between the prediction q̂ and ground truth q. Recall at threshold T is the percentage of evaluated frames satisfying:
A missing or invalid prediction is treated as having infinite error, so it remains in the evaluation denominator and is not recalled at any finite threshold.
In our leaderboard, we report:
To measure the combined geometric effect of the predicted camera parameters, we use a fixed set of 100,000 3D points sampled from high-resolution ETH3D scenes to represent naturally occurring indoor and outdoor structure.
For each benchmark frame, we project the evaluated 3D points using both the predicted and ground truth camera intrinsics. End point error (EPE) is the L2 pixel distance between each corresponding pair of projected locations. EPE recall at threshold T is the percentage of evaluated frame-point pairs whose EPE is strictly below T. If a method does not produce valid camera intrinsics for a frame, its EPE is treated as infinite.
The ground truth intrinsics for InFlux and InFlux++ Real are obtained by interpolation over lens lookup tables (LUTs). To avoid evaluating frames whose ground truth may be unreliable because of LUT interpolation, we use LUT-reliable EPE recall.
For each LUT experiment, we perform leave-one-out (LOO) validation: we hold out its measured intrinsics, interpolate them from the remaining LUT experiments, and compare the interpolated and measured intrinsics using EPE recall at threshold T. When computing LUT-reliable EPE recall at T pixels, we include only frames whose enclosing LUT region has LOO EPE recall at T pixels of at least 0.95 at every vertex.
Because LUT reliability is determined separately for each threshold, the set of eligible frames may differ across thresholds. In our leaderboard, we report LUT-reliable EPE recall at 10, 50, and 300 pixels. See the InFlux++ paper for the full LUT-reliability procedure.
Follow the full submission instructions to install the command-line tools, generate a complete submission template, fill it with predictions, and upload the completed JSON file.
At a high level, influx-generate-sample creates a complete
template for the selected evaluation target, and influx-upload
validates and uploads the completed submission. The benchmark currently
supports the radial-tangential (rad-tan) and Mei
(mei) camera models.
The GitHub guide contains the complete JSON schema, required parameters, command options, validation rules, and troubleshooting guidance.
The upload process verifies ownership through the supplied email address and assigns a submission UUID. Keep this UUID: it permanently identifies the submission and is useful for result tracking, support requests, and later leaderboard changes.
Evaluation results are typically returned by email within a few hours. Queue load or ground truth cache staging may increase the turnaround time.
Results are private by default. After evaluation, the original submitter can use the command-line tools to publish the result, hide a previously published result, or update its public display name, publication information, and code link.
Each email address is limited to three successfully queued submissions within a rolling seven-day period.
If no result email arrives within one to two days, check the spam or junk folder and then contact influxbenchmark@gmail.com. Include the submission UUID, method name, benchmark version, and submitter email.
