AeroBelief: Budgeted First-Action Verification for Evidence-Guided Aerial Sensing
Organizations: Harbin Institute of Technology, China · The Chinese University of Hong Kong, China
Abstract
Aerial sensing must combine uncertain overhead context with current observations while budgeting prediction, movement and image processing. AeroBelief uses the next executable action as the unit of verification against reference scores. It groups candidate views by first action, scans them with an inexpensive proposer, and selectively queries a structured world model (WM). Calibrated group bounds support strict commitment; a six-query controller connects selective prediction with belief updates, acceptance checks and scheduled sensing. Across four layout families, 40 worlds and 960 closed-loop executions, the AeroBelief controller completes 75.42% of tasks, compared with 73.33% for exhaustive WM scoring and 65.83% for line-of-sight geometry, using 91.81% fewer reference-candidate evaluations than exhaustive scoring. In fixed-reference replay with a six-query cap, proposer-scaled anchored bounds raised first-action certificate availability from 36.25% to 54.69% and reduced mean queries per state by 13.54% relative to unit-scale anchored bounds. A complete ground-hosted pipeline processes 2,685 self-collected 20-megapixel unmanned aerial vehicle (UAV) images from seven DJI Mavic 3M routes. Context-guided admission and feedback retain all eight independently supported target confirmations with 76.05% fewer processed images and 70.02% less accumulated active processing service than uncapped all-frame processing; feedback saves another 5.17% of images relative to static context. The studies connect budgeted action selection with evidence-driven processing of recorded flight imagery.
Figures & tables
| Study | Population | Endpoint |
|---|---|---|
| Noisy-sensor controls | 4,320 matched tasks per policy; six policies | Context, negative evidence and qualified confirmation. |
| Empirical-noise transfer | 216 worlds; 6,480 evaluations | Development-route likelihoods and prior mismatch. |
| Crossed verification | 2,160 keys; 21,600 trajectories | Action grouping, refinement and budget frontier. |
| Independent calibration | 3,240 calibration / 6,480 test worlds | Simultaneous coverage and usable decisions. |
| Four-family control | 40 worlds; 960 executions; four arms | Completion, reference work and paired mission costs. |
| Two-regime policy control | 1,440 six-policy + 240 aligned Student executions; 1,680 total | Completion across difficulty regimes and proposer-only controls. |
| Rule | Available tasks | Successful tasks |
|---|---|---|
| Candidate certificate | 26 | 26 |
| Candidate + query refinement | 211 | 211 |
| First-action certificate | 153 | 153 |
| First-action + query refinement | 453 | 450 |
| Policy/access | Time (ms) | WM candidates |
|---|---|---|
| Full, serial | 41.683 | 29.43 |
| Full, native batch | 34.437 | 29.43 |
| Top-6, native batch | 33.837 | 3.70 |
| D6, serial | 35.074 | 2.60 |
| Aligned Student, batch | 34.252 | 0 |
Appendix figures & tables64 assets
Supplementary material from the paper’s appendix.
Appendix
| Contract | Population | Measured finding | Interpretation |
|---|---|---|---|
| Analytic action grouping | 2,160 matched task keys | At , grouping plus refinement makes 453 tasks available and completes 450. | A first action can be resolved without a unique best continuation. |
| Engineering D6 control | 40 worlds; 240 starts/arm | D6/F/ /G complete 181/176/173/158. D6/F candidate work is 14,453/176,451. | Descriptive four-family completion and total logical work. |
| Aligned-Student cost | 229 common-valid starts | Simulated time and energy decrease 7.42% and 7.52%; paired cost intervals exclude zero. | Includes timeouts, excludes unavailable initialization telemetry. |
| Serial CPU interface | 120 states; three repeats | Full/D6 total latency 41.683/35.074 ms. | 15.86% less time when reference access is serial. |
| Native-batch CPU interface | Same 120 states | Full/D6 total latency 34.437/35.074 ms. | D6 takes 1.85% more time with native batching. |
| Reference-interval replay | 10,459 historical states | Submitted disagreements 33/2,309 to 0/1,446; queries 10,116 to 28,922. | A reliability–availability–work trade-off on state replay. |
| Component | Included inputs | External or unresolved inputs |
|---|---|---|
| Simulation | Trial summaries, initial-cohort state ledger, selected historical archives and Student terminal trajectories | Two full four-family raw volumes are externally indexed, including the long/oblique extension’s full decision/array streams. |
| Image processing/review | Detection, schedule, judgment and coordinate exports; some image/query assets | The complete 2,685-photograph collection and review-image sources are externally indexed; included subsets do not supply the full pixel input. |
| Rendered physical evaluation | Scalar candidate labels, fixed choices and progress mappings | Full render masks, complete belief/history, and 240-state candidate probabilities are not all included. |
| Calibration and parity | Fitted maps, probability summaries and parity flags | Some candidate-level fitting arrays are missing; raw reference-calibration world maxima and underlying historical CPU selections are externally indexed. |
| Reference-state replay | Query traces and fixed-score banks | Backbone training-world membership is unknown. |
| Arm | True | False | Unconfirmed | Observations | Motions | Total cost |
|---|---|---|---|---|---|---|
| full | 3629 | 29 | 662 | 25.610 | 24.610 | 31.013 |
| no_context | 1416 | 17 | 2887 | 47.342 | 46.342 | 58.177 |
| frozen | 111 | 1201 | 3008 | 48.866 | 47.866 | 60.082 |
| naive_zero | 3316 | 24 | 980 | 27.637 | 26.637 | 33.547 |
| single_hit | 2733 | 1314 | 273 | 17.770 | 16.770 | 21.213 |
| mass | 3730 | 38 | 552 | 25.136 | 24.136 | 30.420 |
| Contrast | Difference(pp) | Pointwise 95% CI | Gains/losses | |
|---|---|---|---|---|
| Full–no context: true | 51.227 | [49.537,52.894] | 2331/118 | -532.436 |
| Full–frozen negative: true | 81.435 | [80.255,82.593] | 3521/3 | -1050.064 |
| Full–naive zero: true | 7.245 | [6.227,8.264] | 423/110 | -43.538 |
| Single hit–full: false | 29.745 | [28.403,31.134] | 1285/0 | -386.221 |
| Study | Controller | |||
|---|---|---|---|---|
| Study I | IG/coverage | 77.03 (0) | 72.74 (9) | 65.16 (0) |
| Study I | Mass | 83.51 (0) | 78.30 (9) | 71.70 (0) |
| Study I | No context | 32.47 (0) | 30.56 (14) | 27.08 (1) |
| Study I | Frozen | 2.20 (1) | 2.26 (14) | 2.26 (1) |
| Study I | Naive zero | 65.10 (0) | 64.47 (9) | 61.00 (0) |
| Study I | Single hit | 77.26 (15) | 73.67 (13) | 66.03 (0) |
| Controller | Absent false/864 | Fault false/128 | Fault unconf./128 | Fault obs. | Fault cost |
|---|---|---|---|---|---|
| IG/coverage | 1 | 125 | 3 | 3.656 | 3.570 |
| Mass | 1 | 125 | 3 | 3.648 | 3.561 |
| HMM | 1 | 58 | 70 | 31.766 | 38.707 |
| Calibrated power | 1 | 3 | 125 | 54.875 | 67.594 |
| Revisit Mass | 1 | 125 | 3 | 3.648 | 3.561 |
| Revisit HMM | 1 | 57 | 71 | 32.141 | 39.176 |
| Band | Method | Available | Cov (%) | Queries/world | Mismatch | Positive gap |
|---|---|---|---|---|---|---|
| Transported 9 | C6 | 740 | 99.89 | 5.492 | 0 | 0 |
| Transported 9 | C+Q6 | 2,462 | 99.89 | 4.748 | 1 | 0 |
| Transported 9 | A6 | 2,172 | 99.89 | 4.312 | 0 | 0 |
| Transported 9 | A+Q6 | 3,566 | 99.89 | 3.548 | 1 | 0 |
| Transported 9 | A+Q12 | 4,151 | 99.89 | 5.993 | 1 | 0 |
| Transported 9 | A+Q24 | 4,994 | 99.89 | 9.598 | 1 | 0 |
| Arm | Available | Success | Abstain | Canonical | Positive gap | Queries/task |
|---|---|---|---|---|---|---|
| C6 | 26 | 26 | 2134 | 1 | 0 | 6.081 |
| C+Q6 | 211 | 211 | 1949 | 0 | 0 | 8.588 |
| A6 | 153 | 153 | 2007 | 1 | 0 | 6.149 |
| A+Q0 | 78 | 78 | 2082 | 0 | 0 | 0.000 |
| A+Q3 | 293 | 293 | 1867 | 1 | 0 | 4.258 |
| A+Q6 | 453 | 450 | 1707 | 1 | 0 | 9.605 |
| Endpoint | Candidate certificate | First-action certificate |
|---|---|---|
| Reference certificate passes | 5,904 (32.56%) | 7,914 (43.64%) |
| Guard allows model action | 2,795 (15.41%) | 2,961 (16.33%) |
| Allowed action differs from NoWorld | 2,424 | 2,559 |
| Reference-action discrepancies after Guard | 19/2,795 | 38/2,961 |
| New logical queries | 23,035 | 14,321 |
| Cap hits / entered states | 2,405/8,238 | 432/8,238 |
| Arm | Success | Run queries | Queries/entered | Replay queries | Certificates | Permitted |
|---|---|---|---|---|---|---|
| CP | 91/120 | 2662 | 1.755 | 10352 | 5777 | 2318 |
| CU | 93/120 | 2618 | 1.758 | 10329 | 5793 | 2318 |
| AP | 95/120 | 2395 | 1.615 | 9902 | 5789 | 2302 |
| AU | 95/120 | 2441 | 1.634 | 9871 | 5796 | 2302 |
| Representation | Learner | Agree./320 | RMSE | pp [95% CI] | /5 |
|---|---|---|---|---|---|
| Shared | Student | 299 | 0.078214 | – | 5 |
| Raw+S | TabM-small2025 | 296 | 0.072965 | -0.938 [-2.500,+0.625] | 1 |
| Raw+S | MLP | 300 | 0.077809 | +0.312 [-1.562,+2.500] | 0 |
| Raw+S | LightGBM | 299 | 0.073037 | +0.000 [-1.562,+1.875] | 0 |
| Raw+S | XGBoost | 299 | 0.074822 | +0.000 [-1.250,+1.250] | 0 |
| Raw+S | ExtraTrees | 302 | 0.072326 | +0.938 [-0.938,+2.812] | 0 |
| Bounds | Schedule | Certified/320 | Mean queries | Mean calls | Ref. errors |
| Absolute | action round robin | 87 | 4.8750 | 4.8750 | 0 |
| Absolute | leader rival | 110 | 4.7906 | 4.7906 | 0 |
| Absolute | leader rival round2 | 108 | 5.0250 | 2.5125 | 0 |
| Absolute | optimistic leader | 110 | 4.7906 | 4.7906 | 0 |
| Absolute | random | 5 | 5.9469 | 5.9469 | 0 |
| Absolute | topk | 103 | 4.8594 | 4.8594 | 0 |
| Bound | Cap | Certified | Queries |
|---|---|---|---|
| Absolute | 6 | 110 | 4.7906 |
| Relative | 6 | 116 | 4.7313 |
| Normalized | 6 | 175 | 4.0906 |
| Absolute | 12 | 118 | 8.6281 |
| Relative | 12 | 124 | 8.5000 |
| Normalized | 12 | 220 | 6.3344 |
| Population | D6 | Full F | Aligned Student | Geometry G |
|---|---|---|---|---|
| Layout perturbations | 42/60 | 43/60 | 42/60 | 34/60 |
| Longitudinal/side | 37/60 | 30/60 | 31/60 | 26/60 |
| Transverse barriers | 57/60 | 57/60 | 58/60 | 56/60 |
| Long/oblique extension | 45/60 | 46/60 | 42/60 | 42/60 |
| Initial three families | 136/180 | 130/180 | 131/180 | 116/180 |
| All four families | 181/240 | 176/240 | 173/240 | 158/240 |
| Population | Starts/worlds | D6/Student time (s) | Time difference 95% | D6/Student energy (J) | Energy difference 95% |
|---|---|---|---|---|---|
| Joint success | 147/39 | 45.34/50.08 | [ , ] | 3,169.3/3,507.2 | [ , ] |
| Common valid | 229/40 | 68.53/74.02 | [ , ] | 4,790.2/5,179.6 | [ , ] |
| Population | Worlds | Tasks/arm | A | B | C | D |
|---|---|---|---|---|---|---|
| First batch | 20 | 120 | 93 | 34 | 105 | 107 |
| Second batch | 20 | 120 | 73 | 26 | 97 | 93 |
| Pooled descriptive | 40 | 240 | 166 | 60 | 202 | 200 |
| Medium | 15 | 90 | 87 | 40 | 87 | 87 |
| Hard | 20 | 120 | 74 | 11 | 93 | 94 |
| Very hard | 5 | 30 | 5 | 9 | 22 | 19 |
| Population | Contrast | Difference | 95% interval | World sign-flip |
|---|---|---|---|---|
| First batch | C–A | 10.00 | [3.33,18.33] | 0.0625 |
| First batch | D–A | 11.67 | [3.33,22.50] | 0.0625 |
| First batch | D–C | 1.67 | [ ,5.00] | 0.6250 |
| Second batch | C–A | 20.00 | [6.67,34.17] | 0.0156 |
| Second batch | D–A | 16.67 | [4.17,30.83] | 0.0391 |
| Second batch | D–C | [ ,1.67] | 0.4063 |
| Arm | Success | Timeout | Collision | Initialization error | Policy error |
|---|---|---|---|---|---|
| A | 166 | 68 | 0 | 6 | 0 |
| B | 60 | 118 | 19 | 7 | 36 |
| C | 202 | 31 | 0 | 7 | 0 |
| D | 200 | 34 | 0 | 6 | 0 |
| Policy | Success | Rate (%) | Timeout | Initialization | Queries |
|---|---|---|---|---|---|
| A | 67/120 | 55.83 | 51 | 2 | 0 |
| G | 76/120 | 63.33 | 42 | 2 | 0 |
| S | 92/120 | 76.67 | 25 | 3 | 0 |
| F | 90/120 | 75.00 | 28 | 2 | 65,599 |
| T6 | 90/120 | 75.00 | 27 | 3 | 8,754 |
| D6 | 97/120 | 80.83 | 21 | 2 | 5,739 |
| Contrast | Rescues | Harms | Difference (pp) | 95% interval |
|---|---|---|---|---|
| D6 vs A | 35 | 5 | +25.00 | [11.67, 40.00] |
| D6 vs G | 26 | 5 | +17.50 | [7.50, 28.33] |
| D6 vs S | 14 | 9 | +4.17 | [-5.83, 14.17] |
| D6 vs F | 11 | 4 | +5.83 | [0.00, 12.50] |
| D6 vs T6 | 12 | 5 | +5.83 | [-1.67, 13.33] |
| Arm | Regime 1 | Regime 2 | Combined | Init. errors | Queries | Backend calls |
|---|---|---|---|---|---|---|
| A | 67/120 | 46/120 | 113/240 | 7 | 0 | 0 |
| G | 76/120 | 46/120 | 122/240 | 7 | 0 | 0 |
| S | 92/120 | 80/120 | 172/240 | 10 | 0 | 0 |
| F | 90/120 | 77/120 | 167/240 | 7 | 185,418 | 4,206 |
| T6 | 90/120 | 77/120 | 167/240 | 7 | 25,098 | 4,183 |
| D6 | 97/120 | 76/120 | 173/240 | 7 | 16,158 | 16,158 |
| Contrast | Rescues | Harms | Gain(pp) | 95% interval |
|---|---|---|---|---|
| D6-A | 76 | 16 | +25.00 | [17.08,33.33] |
| D6-G | 66 | 15 | +21.25 | [13.75,28.75] |
| D6-S | 38 | 37 | +0.42 | [-6.25,7.08] |
| D6-F | 29 | 23 | +2.50 | [-2.08,7.08] |
| D6-T6 | 29 | 23 | +2.50 | [-2.92,7.92] |
| Arm | Layout I | Layout II | All | Timeout | Init. error | Actual WM queries |
|---|---|---|---|---|---|---|
| A | 67/120 | 46/120 | 113/240 | 120 | 7 | 0 |
| G | 76/120 | 46/120 | 122/240 | 111 | 7 | 0 |
| S | 92/120 | 80/120 | 172/240 | 58 | 10 | 0 |
| 91/120 | 61/120 | 152/240 | 79 | 9 | 0 | |
| F | 90/120 | 77/120 | 167/240 | 66 | 7 | 185,418 |
| T6 | 90/120 | 77/120 | 167/240 | 66 | 7 | 25,098 |
| Policy | Success | Hard/36 | Very hard/24 | Decisions | Total queries | Entered mean |
|---|---|---|---|---|---|---|
| A | 24/60 | 23 | 1 | 1,710 | 0 | – |
| F | 43/60 | 30 | 13 | 1,464 | 40,617 | 45.231 |
| T6 | 36/60 | 26 | 10 | 1,576 | 6,054 | 6.000 |
| D2 | 43/60 | 29 | 14 | 1,480 | 1,479 | 1.585 |
| D6 | 42/60 | 26 | 16 | 1,482 | 3,295 | 3.719 |
| D12 | 43/60 | 27 | 16 | 1,501 | 6,609 | 7.114 |
| Contrast | Rescues | Harms | Difference (pp) | 95% interval (pp) |
|---|---|---|---|---|
| F–A | 23 | 4 | 31.67 | [11.67,53.33] |
| T6–A | 16 | 4 | 20.00 | [3.29,36.67] |
| D2–A | 22 | 3 | 31.67 | [11.67,51.67] |
| D6–A | 21 | 3 | 30.00 | [8.33,51.67] |
| D12–A | 23 | 4 | 31.67 | [5.00,56.67] |
| D6–T6 | 9 | 3 | 10.00 | [ ,25.00] |
| A | F | T6 | D2 | D6 | D12 | |
|---|---|---|---|---|---|---|
| Task wall time (s) | 103.89 | 97.61 | 100.87 | 96.56 | 97.42 | 99.32 |
| Simulated battery (J) | 5,585.04 | 5,149.10 | 5,363.50 | 5,131.83 | 5,086.20 | 5,210.81 |
| Split | Worlds | States | Candidates | Pixel-positive | All zero | Subthreshold |
|---|---|---|---|---|---|---|
| Calibration | 80 | 480 | 23,588 | 139 | 324 | 17 |
| Fixed test | 40 | 240 | 11,884 | 66 | 164 | 10 |
| Visited | 39 | 39 | 1,981 | 6 | 32 | 1 |
| Calibration scope | ||
|---|---|---|
| All states / all candidates | 2.3874 | 6.3777 |
| D Guard accepted / all candidates | 1.8825 | 5.8382 |
| Full-WM Guard accepted / all candidates | 1.5046 | 5.8382 |
| D accepted / selected candidate only | 1.0131 | 5.7313 |
| Measure | Value |
|---|---|
| Student–WM mean absolute error | 0.02137 |
| WM–rendered log-pixel mean absolute error | 0.35586 |
| Probability-head Brier / absent baseline | 0.02972 / 0.01717 |
| Probability-head area under ROC / average precision | 0.4691 / 0.01791 |
| Probability threshold 0.25: TP / FP / FN / TN | 13 / 378 / 173 / 10,271 |
| 35-pixel prediction: TP / FP / FN / TN | 3 / 166 / 183 / 10,483 |
| Quantity | Count |
|---|---|
| True positive | 3 |
| False positive | 166 |
| False negative | 183 |
| True negative | 10483 |
| Predicted positive, actual zero | 150 |
| Actual-zero FP in all-zero states | 123 |
| Rule | Best group /52 | Group /52 | Best group /60 | Distance-best /176 | Mean progress(m) | Moving states |
|---|---|---|---|---|---|---|
| LOS geometry (G) | 14 | 24 | 15 | 27 | 0.124871 | 84 |
| NoWorld (A) | 19 | 30 | 20 | 28 | 0.142363 | 82 |
| Full + Guard | 21 | 32 | 22 | 39 | 0.235326 | 99 |
| Top-6 | 21 | 32 | 22 | 38 | 0.225961 | 98 |
| D6 | 21 | 32 | 22 | 38 | 0.218087 | 98 |
| Raw WM | 14 | 28 | 16 | 33 | 149 |
| Bank / wrapper | Horizon | Nonzero states | Correct scores | Geometry only | Shuffled scores |
|---|---|---|---|---|---|
| 236 / no progress | Endpoint | 60 | 11.72 | 12.01 | 13.52 |
| 236 / no progress | First step | 25 | 55.18 | 47.83 | 48.59 |
| 236 / no progress | Group opportunity | 60 | 51.16 | 45.41 | 40.06 |
| 236 / progress | Endpoint | 60 | 17.17 | 17.65 | 15.52 |
| 236 / progress | First step | 25 | 47.83 | 43.83 | 44.28 |
| 236 / progress | Group opportunity | 60 | 57.79 | 53.93 | 45.07 |
| Cell | Total (ms) | Logical WM candidates | Backend calls | Model rows | Model accepts |
|---|---|---|---|---|---|
| F, cap 1 | 41.683 | 29.43 | 29.43 | 206.03 | 20/120 |
| F, cap 2 | 38.025 | 29.43 | 14.97 | 206.03 | 20/120 |
| F, cap 6 | 36.025 | 29.43 | 5.21 | 206.03 | 20/120 |
| F, cap 12 | 34.980 | 29.43 | 2.89 | 206.03 | 20/120 |
| F, all | 34.437 | 29.43 | 0.62 | 206.03 | 20/120 |
| Top 6, cap 1 | 35.082 | 3.70 | 3.70 | 25.90 | 20/120 |
| Policy | Access | Time(ms) | Queries | Invocations | Seven-row inputs |
|---|---|---|---|---|---|
| D6 | Batch | 37.018 | 2.600 | 2.600 | 18.200 |
| D6 | Per candidate | 36.911 | 2.600 | 2.600 | 18.200 |
| F | Batch | 36.028 | 29.433 | 0.617 | 206.033 |
| F | Per candidate | 43.958 | 29.433 | 29.433 | 206.033 |
| T6 | Batch | 35.964 | 3.700 | 0.617 | 25.900 |
| T6 | Per candidate | 36.541 | 3.700 | 3.700 | 25.900 |
| Path | Mean ms | P95 ms | P99 ms | Queries/state | Feature rows/state |
|---|---|---|---|---|---|
| Full, reuse off | 180.136 | 220.598 | 236.912 | 48.366 | 338.559 |
| Full, reuse on | 35.784 | 42.983 | 46.844 | 48.366 | 48.366 |
| Selective, reuse off | 48.800 | 74.867 | 82.262 | 2.538 | 63.591 |
| Selective, reuse on | 36.382 | 43.956 | 46.433 | 2.538 | 48.366 |
| Gated full, reuse on | 34.509 | 41.479 | 43.534 | 29.072 | 48.366 |
| Quota | Full (ms) | 20,777 (ms) | 14,121 (ms) |
|---|---|---|---|
| Normal | 33.286 | 33.627 | 33.861 |
| 50% | 62.477 | 64.719 | 64.295 |
| 25% | 127.558 | 132.743 | 129.004 |
| 10% | 306.972 | 321.743 | 316.547 |
| 5% | 618.400 | 628.366 | 633.365 |
| 2% | 1,019.425 | 1,385.534 | 1,222.096 |
| Setting | Runs | Successes | Injected | Rejected |
|---|---|---|---|---|
| baseline normal | 8 | 5 | 0 | 0 |
| hardened normal | 8 | 5 | 0 | 0 |
| cache content | 8 | 7 | 8 | 8 |
| model nonfinite | 8 | 8 | 8 | 8 |
| response identity | 8 | 6 | 8 | 8 |
| query timeout | 8 | 5 | 8 | 8 |
| Rule | Task R@1(%) | Mean images |
|---|---|---|
| Fixed 1 | 78.996 | 1.000 |
| Fixed 2 | 81.338 | 2.000 |
| Fixed 3 | 82.394 | 3.000 |
| AeroBelief | 82.059 | 2.600 |
| Frozen margin | 81.390 | 2.004 |
| Frozen entropy | 81.725 | 2.560 |
| Rule | Mean time(ms) | Mean images |
|---|---|---|
| Fixed 1 | 139.46 | 1.000 |
| Fixed 2 | 279.96 | 2.000 |
| Fixed 3 | 418.77 | 3.000 |
| AeroBelief | 365.22 | 2.603 |
| Dev. margin | 415.21 | 2.967 |
| Dev. entropy | 419.31 | 3.000 |
| Encoder | Rule | R@1(%) | Mean images |
|---|---|---|---|
| Frozen | Fixed 1 | 35.235 | 1.000 |
| Frozen | Fixed 2 | 44.907 | 2.000 |
| Frozen | Fixed 3 | 52.725 | 3.000 |
| Frozen | AeroBelief | 52.183 | 2.844 |
| Frozen | Dev. margin | 52.725 | 2.991 |
| Frozen | Dev. entropy | 52.725 | 3.000 |
| Method | Jobs | Target confirmed | Two-full support | Images | Service (s) | Max queue |
|---|---|---|---|---|---|---|
| AllFrames (uncapped) | 8 | 8/8 | 8/8 | 1,762 | 4,987.52 | 244 |
| Static context | 8 | 8/8 | 8/8 | 445 | 1,610.00 | 15 |
| Combined feedback | 8 | 8/8 | 8/8 | 422 | 1,495.23 | 13 |
| Fixed interval | 8 | 4/8 | 3/8 | 603 | 2,130.76 | 9 |
| Method | Target confirmed/8 | Strict full/8 | False confirmed | Images |
|---|---|---|---|---|
| AllFrames | 8 | 8 | 0 | 1,762 |
| Static context | 8 | 8 | 0 | 445 |
| Positive only | 8 | 8 | 0 | 434 |
| Negative only | 8 | 8 | 0 | 428 |
| Combined | 8 | 8 | 0 | 422 |
| One-shot reserve variant | 8 | 8 | 0 | 422 |
| Policy | Image events | Full support | Visible support |
|---|---|---|---|
| Belief | 434 | 8/8 | 8/8 |
| Belief + negative | 422 | 8/8 | 8/8 |
| Static prior | 445 | 8/8 | 8/8 |
| No context | 568 | 5/8 | 6/8 |
| Uniform | 603 | 3/8 | 4/8 |
| Novelty | 526 | 5/8 | 5/8 |
| Task | Frames | Eligible positives | Center inside | Strict overlap |
|---|---|---|---|---|
| D02 old | 356 | 5 | 5 | 5 |
| D02 new | 362 | 8 | 3 | 3 |
| QX WL2 | 678 | 9 | 9 | 9 |
| e0928 | 422 | 7 | 7 | 7 |
| G01 target 1 | 257 | 3 | 3 | 3 |
| G01 target 2 | 257 | 5 | 4 | 3 |
| Cases | Policy | Events | Supported | Pending |
|---|---|---|---|---|
| Development | All frames | 940 | 4/4 | 0 |
| Development | Causal, no context | 361 | 2/4 | 1 |
| Development | Causal context | 237 | 4/4 | 0 |
| Development | Causal context + diagnostic negatives | 227 | 4/4 | 0 |
| Additional tasks | All frames | 822 | 4/4 | 0 |
| Additional tasks | Causal, no context | 207 | 2/4 | 1 |
| Policy | Events | Supported | Time (s) | Pass range (s) | Peak RSS (MiB) |
|---|---|---|---|---|---|
| All frames | 822 | 4/4 | 659.09 | 650.92–667.25 | 146.57 |
| No context, recorded horizon | 356 | 3/4 | 281.86 | 271.59–292.13 | 145.69 |
| Context, recorded horizon | 243 | 4/4 | 191.82 | 191.76–191.88 | 145.52 |
| No context, causal | 207 | 2/4 | 170.87 | 170.60–171.14 | 145.20 |
| Context, causal | 197 | 4/4 | 159.53 | 157.72–161.34 | 145.03 |
| Configuration | Confirmed routes | Total processed frames |
|---|---|---|
| All frames, no early stop | 4/4 | 1,818 |
| All frames, one positive | 4/4 | 813 |
| All frames, two positives | 4/4 | 976 |
| Uniform 120-frame cap | 1/4 | 458 |
| Every two, phase 0 / phase 1 | 4/4; 4/4 | 493; 491 |
| Every five, phases 0/1/2/3/4 | 0/4; 0/4; 1/4; 1/4; 0/4 | 366; 365; 262; 347; 362 |
| Route | Frames | All-frame processed | Context processed | All-frame confirm (s) | Context delay (s) |
|---|---|---|---|---|---|
| D02-old | 356 | 248 | 76 | 1,112 | 0 |
| D02-new | 362 | 219 | 64 | 981 | 0 |
| QX-WL2 | 678 | 169 | 19 | 336 | 6 |
| e0928 | 422 | 340 | 63 | 677 | 0 |
| Total | 1,818 | 976 | 222 | – | – |
UAV Remote Sensing
Unmanned aerial vehicle (UAV) remote sensing uses drone-borne imagery and sensors to map, measure, and monitor Earth's surface.
See all 80 papers