|
| 1 | +# DVIDIA image-motion alpha — 10 October 2026 |
| 2 | + |
| 3 | +**Six authored fixtures processed successfully; real-world accuracy and live |
| 4 | +robot latency remain unmeasured.** This evaluates sparse tracking of a manually |
| 5 | +seeded image region and passive constant-velocity forecasting. It evaluates |
| 6 | +neither an object detector nor a trained interaction or robot-control model. |
| 7 | + |
| 8 | +The [frozen protocol](protocol.json) and [aggregate results](result.json) |
| 9 | +preserve parameters, input/reference hashes, software, attempted counts, |
| 10 | +conditional errors, abstentions and timing scope. They contain no input footage, |
| 11 | +model weights, private data or credentials. The generated fixtures and these |
| 12 | +DVIDIA artifacts are MIT licensed. The existing project notices continue to |
| 13 | +apply to upstream dependencies. |
| 14 | + |
| 15 | +## Frozen inputs and causal separation |
| 16 | + |
| 17 | +Each case is a four-second, 640 × 360, 30 fps video: 120 source frames, 720 frames |
| 18 | +and 24 seconds across the batch. One 80 × 64 pixel textured rectangle translates |
| 19 | +over a uniform background. The seed is manually authored inside that rectangle |
| 20 | +on the first frame. Its center is the image anchor; it is not a measured physical |
| 21 | +centroid. The low-texture negative uses an untextured rectangle. The occlusion |
| 22 | +case removes the rectangle for frames 45–64 while its authored trajectory |
| 23 | +continues. |
| 24 | + |
| 25 | +The deterministic generator freezes trajectories, encoded source hashes and |
| 26 | +actual frame presentation timestamps before tracking. Reference timestamps come |
| 27 | +from integer video PTS multiplied by its rational time base. The tracker receives |
| 28 | +only the manifest, initial seed and processing cadence. Future reference |
| 29 | +positions are opened by the evaluator after processing and checked against their |
| 30 | +frozen hashes. No future reference, simulator state or video frame enters an |
| 31 | +earlier motion estimate or forecast. |
| 32 | + |
| 33 | +The tracker reports the initial seed center plus accumulated robust feature |
| 34 | +translation. Its current estimate is the observed anchor in this unsmoothed |
| 35 | +baseline. Velocity uses two contiguous actual observations and their actual |
| 36 | +time difference. A lost track stays lost, with no replacement, interpolation or |
| 37 | +automatic reacquisition. Missing evidence becomes an abstention. |
| 38 | + |
| 39 | +## Tracking results |
| 40 | + |
| 41 | +| Authored case | Tracked / visible reference frames | Visible track coverage | Anchor error p95, px | Velocity error p95, px/s | |
| 42 | +| --- | ---: | ---: | ---: | ---: | |
| 43 | +| Stationary | 120 / 120 | 100% | 0 | 0 | |
| 44 | +| Constant velocity | 120 / 120 | 100% | 0.00691 | 0.0574 | |
| 45 | +| Stop/start | 120 / 120 | 100% | 0.00698 | 0.0621 | |
| 46 | +| Direction change | 120 / 120 | 100% | 0.00714 | 0.0627 | |
| 47 | +| Brief occlusion | 45 / 100 | 45% | 0.00722 | 0.0582 | |
| 48 | +| Low texture | 0 / 120 | 0% | Unavailable | Unavailable | |
| 49 | + |
| 50 | +All 720 source frames have attempted observation rows. All four textured cases |
| 51 | +without occlusion retained their anchors. The occlusion case lost the track |
| 52 | +when the object disappeared and remained lost after reappearance: 75 lost rows, |
| 53 | +including 55 visible rows. The low-texture case abstained on all 120 frames. |
| 54 | +Neither negative case is a perfect prediction result; the accuracy sample |
| 55 | +counts become zero where no valid anchor/forecast exists. Velocity scores |
| 56 | +exclude the two-frame neighborhood of authored speed discontinuities. |
| 57 | + |
| 58 | +## Forecast results and failure tails |
| 59 | + |
| 60 | +Errors are image-plane endpoint distances in pixels. Hold-current and |
| 61 | +constant-velocity forecasts below share the same evaluable current anchors, |
| 62 | +reference targets and sample counts. Scores are conditional on producing a |
| 63 | +forecast; their answer coverage is reported separately. |
| 64 | + |
| 65 | +| Authored case | 100 ms CV mean / hold mean, px | 500 ms CV mean / hold mean, px | 500 ms CV p95 / worst, px | Forecast samples at 100 / 500 ms | |
| 66 | +| --- | ---: | ---: | ---: | ---: | |
| 67 | +| Stationary | 0 / 0 | 0 / 0 | 0 / 0 | 116 / 104 | |
| 68 | +| Constant velocity | 0.00675 / 9.005 | 0.0252 / 45.005 | 0.0337 / 0.0914 | 116 / 104 | |
| 69 | +| Stop/start | 0.317 / 5.903 | 6.937 / 27.699 | 38.531 / 45.007 | 116 / 104 | |
| 70 | +| Direction change | 0.317 / 8.897 | 6.945 / 41.769 | 59.109 / 89.975 | 116 / 104 | |
| 71 | +| Brief occlusion | 0.00664 / 9.006 | 0.0252 / 45.006 | 0.0342 / 0.0914 | 44 / 44 | |
| 72 | +| Low texture | Unavailable | Unavailable | Unavailable | 0 / 0 | |
| 73 | + |
| 74 | +Every case has 117 available future reference frames at 100 ms and 105 at |
| 75 | +500 ms. Terminal forecasts are censored: three and fifteen rows respectively. |
| 76 | +The first tracked row has no velocity and therefore no forecast. The four |
| 77 | +continuous textured cases answer on 116/117 and 104/105 available targets. |
| 78 | +Occlusion answers on 44/117 and 44/105; low texture answers on zero. The remaining |
| 79 | +rows are explicit abstentions, not zero-error samples. All 250 ms metrics and |
| 80 | +denominators are also retained in the result file. |
| 81 | + |
| 82 | +Current anchors match the same acquisition PTS within a 2 μs numeric |
| 83 | +serialization tolerance. Off-grid observations cannot borrow nearby reference |
| 84 | +truth. Future targets use the nearest actual source frame within half a frame |
| 85 | +(1/60 second), with the earlier frame chosen on an exact tie. The 250 ms horizon |
| 86 | +lands between 30 Hz frames; its reference is therefore approximately 16.67 ms |
| 87 | +from the nominal target. The constant-velocity case consequently has about |
| 88 | +1.49 px mean error there even though its 100/500 ms error is much smaller. |
| 89 | +These are nearest-frame endpoint scores, with offsets disclosed, rather than |
| 90 | +exact-time 250 ms accuracy. No backward interpolation produces new references. |
| 91 | + |
| 92 | +The worst direction-change forecast misses by approximately 90 px at 500 ms. |
| 93 | +Its low whole-case mean does not make it reliable at the turn. Passive |
| 94 | +extrapolation cannot know an unannounced future reversal or contact action. |
| 95 | +Interaction-conditioned models and evaluation slices remain subsequent work. |
| 96 | +The propagated uncertainty radii are heuristic and uncalibrated; their empirical |
| 97 | +coverage and size are recorded without probability claims. |
| 98 | + |
| 99 | +## Processing and reproducibility |
| 100 | + |
| 101 | +Six of six attempted cases completed with no pipeline failures. Processing took |
| 102 | +**1.454 seconds**, plus **1.514 seconds** for fixture generation. The offline |
| 103 | +batch-time/source-duration ratio is **0.0606**. Processing includes intake, |
| 104 | +probing, streaming decode, causal tracking/forecast, writes and scoring; it |
| 105 | +excludes fixture generation and provisioning. The ratio is withheld for any |
| 106 | +failed batch so early failure cannot appear to improve throughput. |
| 107 | + |
| 108 | +This was one run on Darwin 25.5.0, ARM64, ten logical CPUs, Python 3.12.9, |
| 109 | +OpenCV 5.0.0 and NumPy 2.5.3. One OpenCV thread was requested, while OpenCV |
| 110 | +reported ten. Actual peak thread concurrency, enforcement of the requested |
| 111 | +limit and a minimum hardware requirement are unmeasured. |
| 112 | +The process-lifetime resident high-water mark was **83,591,168 bytes |
| 113 | +(79.72 MiB)**, including fixture generation; it is not a per-stage limit. |
| 114 | +No GPU or paid inference API is used in this baseline. A fast batch result does |
| 115 | +not establish live capture-to-action latency, worst-case deadlines, long-running |
| 116 | +stability or robot-control benefit. |
| 117 | + |
| 118 | +From an installed checkout containing the motion alpha, with the observations |
| 119 | +extra and FFmpeg/ffprobe installed: |
| 120 | + |
| 121 | +```sh |
| 122 | +dvidia-motion benchmark --output runs/my-motion-benchmark |
| 123 | +dvidia-motion inspect runs/my-motion-benchmark |
| 124 | +dvidia-motion studio runs/my-motion-benchmark --port 8286 |
| 125 | +``` |
| 126 | + |
| 127 | +Use a new output directory each time. Complete local replay bundles retain |
| 128 | +generated videos, actual-clock frame metadata, causal rows, seed evidence and |
| 129 | +all attempted receipts. This public aggregate is not itself a replay bundle. |
| 130 | +Generator and implementation hashes bind the protocol; encoder/platform |
| 131 | +differences may change encoded-file hashes, so compare declared software and |
| 132 | +parameters before interpreting reproduction differences. |
| 133 | + |
| 134 | +These six cases share one authoring recipe and texture. They are engineering |
| 135 | +checks rather than six independent real-world task attempts or a held-out |
| 136 | +generalization cohort. Next evidence should include unseen permitted footage, |
| 137 | +camera movement, blur, deformation and separate live stream measurements. |
0 commit comments