# Full versus Lite hand models and image preprocessing The saved-image comparison favors the Full palm detector and Full landmark model. Neither contrast normalization nor lens correction improved usable bare-hand detections on the two available bare-hand scenes. This is a small static corpus, not a live tracking pass. ## Method The same saved 640 × 480 grayscale stereo pairs were evaluated with eight configurations: Full or Lite palm/landmark bundles, each with raw input, calibrated lens correction, 1st/99th-percentile contrast normalization, or both corrections. Grayscale-to-RGB replication and MediaPipe's model-required tensor scaling were enabled in every configuration. The main comparison ran on the Samsung SM-F976U using the app's Android MediaPipe Tasks 1.0.0 and OpenCV 4.13.0. Each configuration created fresh CPU VIDEO detectors, with two hands per camera, parallel camera inference, and all detection/presence/tracking thresholds at 0.6. Adaptive input and scheduled one-hand inference were excluded. Each pair received one cold inference, four warmups, and twelve measured repetitions at synthetic 80 ms timestamp intervals. Bare-hand mode was verified; wrist-only mode was false. This was an offline replay with the game idle and tracking cameras disconnected. It loaded the phone's saved verified calibration and never initialized cameras or published replay poses. Raw-image landmarks were undistorted during triangulation; corrected-image landmarks were triangulated without applying distortion correction again. Both paths retained the existing handedness, depth, reprojection, palm geometry, and motion checks. The corpus contains two bare-hand scenes, one no-hand control, and six controller-grip scenes. The central bare-hand scene was also repeated at the end of the phone run. Replaying one pair twelve times does not create twelve independent accuracy samples. Desktop IMAGE and VIDEO tests using MediaPipe 0.10.21/OpenCV 4.11.0 provide a separate cross-check. ## Bare-hand result For capture `64390821246320`, two open hands are visible in both cameras: | Image preprocessing | Full: usable stereo palms per replay | Lite: usable stereo palms per replay | |---|---:|---:| | Raw | 1 of 2 | 0 of 2 | | Lens correction only | 1 of 2 | 0 of 2 | | Contrast normalization only | 1 of 2 | 0 of 2 | | Lens correction + contrast normalization | 1 of 2 | 0 of 2 | The cold frame and all twelve measured frames agreed. Full found one hand in each eye; Lite found no hands with raw, lens-only, or contrast-only input. With both corrections, Lite found one hand in the left camera and none in the right, so it could not reconstruct a stereo position. Capture `72745789186883` contains one open hand, peripheral in the left image. Every configuration missed it in both cameras throughout cold and measured replay. The desktop IMAGE and VIDEO tests agreed with these bare-hand outcomes. The no-hand control `59405518228482` returned zero hands across all configurations. The final repeat of the central scene reproduced the same cold and measured counts in every configuration. All six controller-grip pairs (`110854218431877`, `110858058677761`, `110862058272915`, `110935538464606`, `110939418512521`, `110943418461791`) produced zero accepted stereo palms and zero accepted wrists in all measured phone replays. Some Full configurations produced individual camera detections, but these did not become valid stereo positions. Grip scenes are reported separately from the bare-hand comparison. ## Phone timing These are median milliseconds for a parallel stereo pair on the central scene. They include RGB preparation, inference, and result extraction, but exclude decoding, precomputed lens correction, model initialization, and stereo reconstruction. They are not end-to-end latency. | Model | Preprocessing | First run | Final repeat | |---|---|---:|---:| | Full | Raw | 46.10 | 82.98 | | Full | Lens only | 75.17 | 84.46 | | Full | Contrast only | 85.64 | 82.99 | | Full | Lens + contrast | 84.10 | 80.49 | | Lite | Raw | 43.07 | 40.06 | | Lite | Lens only | 37.29 | 39.70 | | Lite | Contrast only | 38.73 | 42.54 | | Lite | Lens + contrast | 78.78 | 75.34 | Android reported thermal status 2 throughout the saved-pair runs. Configuration order was fixed, and inference work differs when a hand is found. Full/raw's first-versus-last variation also prevents attributing the timing differences to preprocessing alone. Lite's lower timings mostly accompany zero returned hands; its two-camera usable-hand result remains zero. ## What normalization changes The saved calibration uses pinhole radial distortion coefficients (`fisheye: false`). The comparison tests that calibrated lens correction, rather than forcing a fisheye model onto the coefficients. Keeping the original camera matrix and output dimensions narrows the field of view: the left output samples approximately raw x=96–544 and y=56–437, and the right approximately x=93–547 and y=57–440. Some peripheral fingers near raw x=600 disappear from the corrected left image. This is a concrete limitation of the current mapping, although the raw version also failed to detect that hand; the crop alone does not explain all failures. Model tensor normalization was already provided by MediaPipe. The additional normalization tested here changes image contrast. Replicating grayscale into RGB satisfies the input channel contract but cannot recover missing color information. ## Recommendation and limits Use Full with raw grayscale replicated into RGB as the next live baseline, retaining MediaPipe's tensor scaling and calibrated point correction for stereo geometry. Full has the stronger detection result here; extra contrast/lens preprocessing has no demonstrated acceptance gain on these scenes. Broader motion testing is still needed before calling any configuration best overall. Full itself still misses one of two central hands and the peripheral hand. The comparison used a temporary diagnostic build with opt-in preprocessing and offline replay. That comparison code has been removed after selecting the Full/raw baseline; the live path keeps grayscale-to-RGB conversion and calibrated point correction. Build, twenty focused JVM tests, packaged runtime verification, and the installed Android replay were checked for the experiment. Local evidence is under `build/hand-comparison/`: `phone/` contains full per-frame results, thermal snapshots and before/after status; `phone-summary.json` contains compact counts; `desktop/` and `desktop-video/` contain the independent comparisons. The scripts are `compare.py`, `compare_video.py`, `run_phone.py`, and `report_phone.py` in that directory. `input-manifest.json` records all capture, calibration, and model hashes plus runtime versions.