Skip to content
SearchSubscribe
TPL-2026-005·preprint·2026-04-30

Multi-Stage Fallback in Real-Time Computer Vision: A Methodology Study

computer-visionobject-detectionreal-time-inferencepipeline-designmethodology

Abstract

Real-time computer vision systems that must operate under latency constraints face a tension between detection accuracy and inference speed. Single-stage pipelines optimized for peak accuracy often fail the latency budget under adverse conditions (motion blur, occlusion, low contrast), while single-stage pipelines optimized for speed sacrifice recall in those same conditions. Multi-stage fallback architectures address this by chaining detectors of increasing cost and decreasing confidence threshold: a fast primary stage handles the common case, and progressively slower stages activate only when the primary stage produces no confident detection. We present a methodology study of this pattern using simulated and public-benchmark data (COCO val 2017), characterizing the tradeoff space across stage count, confidence threshold, and latency budget. Under a 40 ms end-to-end budget, a three-stage pipeline achieves estimated mean F1 of 0.83 (95% CI: 0.79–0.87, n=500 simulated sequences) versus 0.71 (95% CI: 0.67–0.75) for a single-stage speed-optimized baseline. The improvement comes at a cost: worst-case latency for the three-stage pipeline approaches the budget ceiling in dense-occlusion scenarios. We discuss the design decisions required to implement this pattern safely (threshold calibration, stage activation logic, and graceful degradation when all stages fail) and document the reproducibility parameters for practitioners wishing to evaluate the pattern on their own detection domains.

1. Introduction

Real-time object detection systems face a latency-accuracy tradeoff that has been extensively characterized in the literature. Early single-stage detectors such as YOLO [3] and SSD [8] demonstrated that anchor-based regression networks could operate at frame rates suitable for real-time applications, at the cost of detection recall under small-object and occluded-object conditions. Later transformer-based architectures, beginning with DETR [1] and refined in RT-DETR [2], improved recall under those conditions but at inference costs that, depending on hardware, can approach or exceed practical latency budgets for edge-deployed systems.

The dominant engineering response to this tradeoff has been to select a single detector and tune its confidence threshold to meet a recall target, accepting whatever precision results. A less common but arguably more principled alternative is to chain multiple detectors in a fallback sequence: invoke the fastest detector first, and activate progressively more expensive detectors only when the earlier stages return no confident detection. This pattern was implicit in the cascade classifiers of Viola and Jones [6], where it was motivated by the observation that most image windows are rejected by fast early stages and only a small fraction require the full classifier chain.

Modern detection pipelines have largely abandoned explicit fallback cascades in favor of end-to-end architectures trained jointly. However, in deployment contexts where latency budgets are fixed, hardware is constrained, and no single model achieves satisfactory recall across all operating conditions, the multi-stage fallback pattern remains a practical option. To our knowledge, there is no published methodological treatment of how to design, calibrate, and evaluate such pipelines for general-purpose object detection on modern neural architectures. This study attempts to fill that gap.

We present a simulated evaluation of three-stage fallback pipelines using architectures representative of the current landscape (transformer-based primary, lightweight CNN fallback, region-proposal heuristic), characterize the tradeoff space across confidence threshold and latency budget, and provide reproducibility parameters sufficient for a practitioner to replicate the evaluation on their own detection domain.

2. Background and Related Work

2.1 Detection architectures and the latency-accuracy frontier

Single-stage detectors (YOLO family [3] [7], SSD [8], EfficientDet [12]) predict bounding boxes and class probabilities in a single forward pass, typically achieving 10–30 ms inference on mid-range GPU hardware. Two-stage detectors (Faster R-CNN [4]) use a region proposal network to generate candidate regions before the classification head, improving recall under occlusion at the cost of roughly 2–5× the inference time of comparable single-stage models. Transformer-based detectors (DETR [1], RT-DETR [2]) replace the anchor mechanism entirely with attention over image features, producing state-of-the-art recall on COCO benchmarks [5] while maintaining inference latencies competitive with single-stage models on sufficiently capable hardware.

For edge deployments on CPU-class or constrained-GPU hardware, all of these architectures require tradeoffs. Lightweight variants (MobileNet-backed detectors [11], small YOLO variants) reduce inference cost but sacrifice recall under adverse conditions. The operating regime of interest for this study is one in which no single model achieves both recall >0.85 and latency <25 ms on the target hardware — a regime common in edge-deployed systems operating below GPU-class compute.

2.2 Cascade and fallback patterns

The Viola-Jones cascade [6] demonstrated that a sequence of increasingly complex classifiers, each with a low false negative rate, could achieve high overall recall while keeping average inference cost well below the cost of applying the final-stage classifier to every candidate. The key insight — that the cost of a cascade is dominated by the fraction of inputs that reach the expensive final stage — applies equally to modern fallback pipelines.

Modern neural cascades have been explored in the context of network pruning and early-exit architectures, but these typically involve sharing representations across stages rather than chaining independent detectors. The multi-stage fallback pattern studied here treats each stage as a black-box callable, which has the practical advantage of permitting heterogeneous architectures at each stage (e.g., a transformer primary, a CNN fallback, and a handcrafted heuristic) without requiring joint training.

2.3 Confidence calibration

A multi-stage pipeline depends critically on the confidence scores produced by each detector being well-calibrated: a score of 0.55 should correspond to approximately 55% empirical accuracy on held-out data. Modern neural detectors are not well-calibrated by default; Niculescu-Mizil and Caruana (2005) [10] documented systematic overconfidence in sigmoid-output classifiers, a pattern that extends to detection confidence heads. Platt scaling [9]applied post-hoc to each stage's confidence output is the calibration approach used in this study (see Section 3.3).

3. Methods

3.1 Pipeline architecture

The three-stage pipeline evaluated in this study consists of the stages described in Table 1. Stage 1 is a transformer-based detector representing the high-recall, higher-latency end of the modern detection landscape. Stage 2 is a lightweight CNN representing a speed-optimized single-stage detector. Stage 3 is a classical region-proposal heuristic (background subtraction followed by contour detection) with no learned parameters. The heuristic stage is not a neural detector and produces bounding-box proposals rather than class predictions; its output is assigned a fixed class label (the single target class) and treated as a low-confidence detection.

StageDetector classConfidence thresholdMax stage latency (ms)Activation condition
1 — Primary1Transformer-based0.5522Always (first stage)
2 — Fallback2Lightweight CNN0.4012Stage 1 returns no detection above threshold
3 — Heuristic3Region-proposal heuristic (background subtraction + contour)0.256Stages 1 and 2 both return no detection
Table 1 — Pipeline stage specifications for the three-stage configuration evaluated in this study. Detector class names are illustrative; practitioners should substitute architectures appropriate for their deployment hardware and domain. Confidence threshold and latency budget values are those used in the simulated evaluation; see Section 2.3 for calibration procedure. 'Activation condition' describes when the stage is invoked during runtime.

Total budget across all three stages: 40 ms. Stage 3 budget is fixed; Stages 1 and 2 budgets represent empirical P95 latencies, not hard cutoffs. If Stage 1 or 2 exceeds its budget, the next stage is skipped to preserve the overall deadline.

The activation logic is simple: Stage 1 runs first on every frame. If its maximum output confidence exceeds the calibrated Stage-1 threshold, the detection is accepted and Stages 2 and 3 are skipped. If Stage 1 produces no detection above threshold, Stage 2 is invoked. If Stage 2 also produces no detection above its (lower) threshold, Stage 3 is invoked. If all three stages fail to produce a detection above their respective thresholds, the pipeline emits a “no detection” result for that frame.

3.2 Simulation methodology

Because the goal of this study is methodological rather than domain-specific, we evaluated the pipeline on synthetically generated sequences rather than on a proprietary domain dataset. Synthetic sequences allow controlled manipulation of occlusion fraction and motion blur sigma, isolating the conditions under which fallback stages contribute most to coverage. Each sequence consists of 30 frames with a single synthetic target object (a colored rectangle) placed at random positions against natural-image backgrounds drawn from COCO val 2017 [5]. Three difficulty levels were generated: clear (no occlusion, no blur), partial occlusion (25–50% of the target occluded), and adverse (partial occlusion plus motion blur at 2–4 px Gaussian sigma). Detector responses were simulated from parametric models fit to published per-class precision and recall curves for a transformer-based detector and a compact CNN detector on COCO val 2017; the heuristic stage was modeled from first principles (contour detection precision decreasing monotonically with occlusion and blur).

This simulation approach has limitations that are discussed in Section 7. Practitioners should treat all reported metric values as illustrative of the methodology rather than as ground-truth performance predictions for any specific detector or domain.

3.3 Confidence calibration procedure

Each stage's confidence threshold was calibrated using a held-out calibration split (20% of each difficulty level, held out from the test set). For neural stages, Platt scaling [9] was applied to the raw confidence outputs before threshold selection. Thresholds were selected to maximize F1 on the calibration split. For the heuristic stage, no learning was involved; the threshold was set to the minimum contour area (in pixels) above which detections were accepted.

3.4 Evaluation metrics

We report precision, recall, and F1 at the frame level (a detection is counted as correct if its IoU with the ground-truth box exceeds 0.5 and its class is correct). We also report coverage rate — the fraction of frames for which at least one stage produces a detection above threshold — and mean end-to-end latency including all stage invocations. All metrics are reported as means with 95% confidence intervals; Wilson score intervals for proportions, normal intervals for continuous quantities, with sample sizes noted.

4. Results

4.1 Latency profile across confidence thresholds

Figure 1 shows mean inference latency as a function of the Stage-1 confidence threshold for each of the three stage types evaluated in isolation. The primary transformer detector exhibits near-constant latency across threshold values — consistent with its architecture running a full forward pass regardless of output confidence — while the lightweight CNN and heuristic stages show latency increasing at lower thresholds because more candidate regions require processing. This has an important design implication: in the full pipeline, lowering the Stage-1 threshold reduces the fraction of frames that fall through to Stage 2, but does so by accepting more Stage-1 detections, including false positives. The threshold is not a free parameter — it trades off Stage-1 precision against fallback frequency.

Figure 10.0010.0020.0030.0040.0050.000.200.300.400.500.600.700.80MEAN LATENCY (MS)CONFIDENCE THRESHOLDPrimary (transformer)Fallback (lightweight CNN)Heuristic stage
Figure 1 — Mean inference latency (ms) as a function of confidence threshold for three detector configurations: a transformer-based primary detector, a lightweight convolutional fallback, and a region-proposal heuristic. Simulated to illustrate methodology; latency profile drawn from published inference benchmarks on edge-class hardware (see text). Error bars show ±1 SD across n=100 simulated frames per threshold value. At threshold 0.50 the primary detector handles the large majority of frames within the latency budget; below 0.35 the heuristic stage begins activating frequently, driving average latency upward.

4.2 Per-stage precision, recall, and F1

Figure 2 shows per-stage precision, recall, and F1 for the three-stage pipeline. Stage 1 achieves the highest precision (0.91, 95% CI: 0.87–0.94) but the lowest recall (0.74, 95% CI: 0.70–0.78) across all sequences. Stage 2 reverses this profile, producing higher recall (0.88, 95% CI: 0.84–0.91) at lower precision (0.79, 95% CI: 0.74–0.83). Stage 3 achieves the highest recall (0.93, 95% CI: 0.90–0.96) at the lowest precision (0.61, 95% CI: 0.56–0.66). The pattern is consistent with the architectural tradeoffs described in Section 2.1 and suggests that the three stages span a meaningful range of the precision-recall curve.

Taken together, these per-stage metrics do not directly describe the full pipeline's performance, because the pipeline does not run all stages on every frame. Section 4.4 describes the stage-assignment distribution, which determines how these per-stage numbers combine into an overall pipeline metric.

Figure 20.000.200.400.600.801.00SCORE0.910.740.82Stage 1 — Primary0.790.880.83Stage 2 — Fallback0.610.930.74Stage 3 — HeuristicPrecisionRecallF1
Figure 2 — Per-stage precision, recall, and F1 for the three-stage pipeline evaluated on simulated sequences (n=500). Stage 1 (primary transformer detector) shows high precision but reduced recall under occlusion conditions. Stage 2 (lightweight CNN fallback) recovers a substantial fraction of Stage 1 misses at the cost of increased false positives. Stage 3 (heuristic region proposal) provides a further recall increment with the lowest precision of any stage. Values are means across all sequences; 95% confidence intervals (Wilson score for proportions, n=500) are ±0.03–0.05 for all cells. Simulation methodology described in Section 2.

4.3 Routing and the confusion matrix

Figure 3 shows the routing confusion matrix for the full three-stage pipeline across all simulated sequences (n=500, all difficulty levels pooled). The dominant entry is the (Stage-1 correct, Stage-1 correct) cell (348 sequences), indicating that Stage 1 handles the majority of detections correctly. The next-largest entries are the (Stage-2 correct, Stage-2 correct) cell (61) and the off-diagonal (Stage-1 correct, Stage-2 correct) cell (22), the latter representing cases where Stage 1 produced a false positive that Stage 2 subsequently corrected. The (All stages failed) diagonal cell (6) represents the irreducible miss rate of the full pipeline at the threshold values used.

Notably, no entry above the anti-diagonal is large, which would indicate a stage producing a correct detection only to have a later stage override it with an incorrect one. The pipeline's activation logic — later stages activate only when earlier stages produce no confident detection — prevents this failure mode by construction.

Figure 3PREDICTED CLASSACTUAL CLASSStage-1 correctStage-2 correctStage-3 correctAll stages failedStage-1 correctStage-2 correctStage-3 correctAll stages failed3482246186183391124236
Figure 3 — Confusion matrix for the full three-stage pipeline on simulated sequences (n=500 sequences, integer counts). Rows are ground-truth outcome; columns are pipeline-assigned outcome. "Stage-1 correct" indicates the primary detector produced a confident, correct detection. "Stage-2 correct" and "Stage-3 correct" indicate the fallback stages recovered a detection that Stage 1 missed. "All stages failed" indicates no stage produced a confident detection above threshold. The dominant diagonal indicates correct routing in the majority of sequences. Off-diagonal entries primarily reflect Stage-1 false positives (row: Stage-1 correct, column: Stage-2 correct) — cases where the primary detector was confident but incorrect, and a fallback stage provided a better answer. Simulated to illustrate methodology; class-balance and error rates derived from published COCO val 2017 benchmark characteristics.

4.4 Coverage-latency tradeoff

Figure 4 shows coverage rate as a function of latency budget for one-stage, two-stage, and three-stage configurations. At a 40 ms budget, the three-stage configuration achieves 0.94 coverage versus 0.92 for two-stage and 0.82 for one-stage. The marginal return of the third stage (0.02 coverage increase over two-stage) is smaller than the marginal return of the second stage (0.10 increase over one-stage), suggesting diminishing returns to stage count. The crossover point — above which adding a third stage provides less than 1 pp of coverage gain — occurs at approximately 35 ms budget in these simulations.

Figure 40.500.600.700.800.901.001520253035404550COVERAGE RATELATENCY BUDGET (MS)One-stageTwo-stageThree-stage
Figure 4 — Coverage rate (fraction of sequences producing a confident detection from any stage) as a function of end-to-end latency budget (ms) for one-stage, two-stage, and three-stage pipeline configurations. Simulated to illustrate methodology. Each point is the mean coverage across n=500 sequences at that latency ceiling; the budget is enforced by aborting stage transitions that would exceed it. At a 40 ms budget, the three-stage configuration achieves 0.94 coverage versus 0.82 for one-stage. The marginal return of the third stage relative to the second diminishes above 35 ms, suggesting that for budgets above that threshold, a two-stage pipeline may be sufficient. Error bars (±1 SD) omitted for readability; SD was 0.02–0.04 across all budget levels.

4.5 Stage-assignment by difficulty

Table 2 shows the stage-assignment distribution across difficulty levels. In clear sequences, Stage 1 handles 83% of frames successfully. In adverse sequences, this drops to 41%, with Stages 2 and 3 together handling 50% of frames. The shift in assignment distribution is the operational signature of the fallback pattern working as intended: the pipeline adapts its cost profile to the actual difficulty of each frame, spending more compute where it is needed and less where it is not.

Sequence difficultyStage 1 (fraction)Stage 2 (fraction)Stage 3 (fraction)All failed (fraction)
Clear (n=500)Clear0.830.120.030.02
Partial occlusion (n=500)Partial occlusion0.610.240.090.06
Adverse (n=500)Adverse0.410.330.170.09
Table 2 — Stage-assignment distribution across three sequence-difficulty categories (simulated, n=500 sequences each). 'Clear' sequences have no occlusion and normal contrast. 'Partial occlusion' sequences have 25–50% target occlusion. 'Adverse' sequences combine partial occlusion with motion blur (sigma > 2 px). Values are fractions of sequences assigned to each stage as the primary successful detection source. Fractions within each difficulty category sum to 1.00 (rounding). The shift from Stage-1 dominance in clear sequences to Stage-2/3 dominance in adverse sequences characterizes the operational regime where fallback stages contribute meaningfully to coverage.

95% confidence intervals (Wilson score) for each fraction are ±0.02–0.04 at n=500. Differences between difficulty categories are statistically significant at p < 0.001 (chi-square test on the 3×4 contingency table, χ²(6) = 184.3).

4.6 Ablation: threshold calibration vs stage count

Figure 5 presents the ablation results. The gap between the three-stage calibrated configuration (F1: 0.83, 95% CI: 0.79–0.87) and the three-stage uncalibrated configuration (F1: 0.77, 95% CI: 0.73–0.81) is 0.06 F1 points, comparable in magnitude to the gap between the two-stage and one-stage configurations (0.79 vs 0.71, difference: 0.08). This suggests that threshold calibration contributes roughly as much to overall performance as adding one additional pipeline stage. Both interventions are worth implementing; calibration is considerably cheaper in engineering effort.

The accuracy baseline (F1: 0.88) represents an upper bound achievable by running the high-accuracy detector on every frame without a latency constraint. The three-stage pipeline's F1 of 0.83 represents approximately 56% of the gap between the speed baseline (0.71) and the accuracy baseline (0.88), under a 40 ms latency budget.

Figure 50.000.200.400.600.801.00MEAN F10.71Speed baseline0.88Accuracy baseline0.79Two-stage0.83Three-stage0.77Three-stage (no cal.)
Figure 5 — Ablation study: mean F1 and mean latency (ms) across pipeline configurations on simulated sequences (n=500 each). "Speed baseline" is a single-stage lightweight detector with no fallback. "Accuracy baseline" is a single-stage high-accuracy detector with no latency constraint. "Two-stage" and "Three-stage" are the multi-stage fallback configurations described in Section 2. "Three-stage (no threshold tuning)" uses default confidence thresholds without per-stage calibration. The gap between calibrated and uncalibrated three-stage configurations (F1: 0.83 vs 0.77) suggests that threshold calibration contributes meaningfully to the observed improvement, independent of the stage-count effect. Simulated to illustrate methodology.

5. Discussion

Three findings from this study warrant discussion.

Fallback stages are most valuable under adverse conditions, not in aggregate. Pooled across all difficulty levels, the three-stage pipeline improves F1 by 0.12 over the speed baseline. But in adverse sequences, the improvement is substantially larger — Stage 1 alone achieves 0.41 coverage in adverse conditions, meaning fallback stages are responsible for recovering more than half of correct detections in that regime. If adverse conditions (occlusion, blur, low contrast) represent a meaningful fraction of real deployment frames, the case for multi-stage fallback is strong. If adverse conditions are rare, a single well-calibrated detector may be sufficient.

Threshold calibration is underappreciated. The 0.06 F1 improvement from calibration (Section 4.6) is achievable with a single Platt-scaling fit on a modest held-out set, requiring no architectural changes and no retraining. In the authors' experience this step is frequently skipped in deployed pipelines — detectors are shipped with default confidence thresholds, which systematically overstate confidence on out-of-distribution inputs. The calibration step described in Section 3.3 takes on the order of 15 minutes of engineering time and is worth doing before evaluating any other pipeline modification.

The “all stages failed” case requires explicit design. Any multi-stage pipeline will occasionally produce no confident detection. The routing matrix (Figure 3) shows 6/500 sequences in the “all stages failed” category under our simulation parameters. In practical deployments, this rate will depend on domain difficulty and threshold settings, and the appropriate response to a failed frame (carry-forward the previous detection, emit a null, trigger an alert) is a design decision that the pipeline cannot make automatically. Systems that do not explicitly handle this case tend to propagate the last confident detection indefinitely, which can produce misleading outputs when the target has genuinely left the scene.

6. Limitations

  • Simulated data. All reported metrics are derived from a simulation parameterized on published COCO benchmarks, not from a real domain deployment. Real detection tasks typically involve richer scene variation, domain shift between training and deployment distributions, and class imbalances not captured by the simulation. The quantitative results should be treated as illustrative of the methodology rather than as performance predictions for any specific system.
  • Single-class setting. The simulation evaluates a single-class detection task. Multi-class pipelines introduce the additional complication that different stages may assign different class labels to the same bounding box, requiring a stage-routing policy that accounts for per-class confidence rather than a single scalar threshold. This complication is not addressed in the present study.
  • Independent stage calibration. Threshold calibration was performed independently for each stage on a held-out calibration split. In deployment, calibration distributions may shift over time or across scene conditions, requiring periodic recalibration. The sensitivity of pipeline performance to calibration drift is not characterized here.
  • No hardware-in-the-loop evaluation. Latency values are derived from published inference benchmarks and parametric simulation, not from measurements on physical hardware. Actual latency will depend on hardware model, driver version, thermal state, and concurrent workloads. The latency budget analysis in Section 4.4 should be re-validated against real hardware before using it to make deployment decisions.
  • Heuristic stage generalization. The Stage-3 heuristic (background subtraction + contour detection) assumes a relatively static background, which is violated in many real deployment settings. In high-motion or outdoor environments, this stage would produce substantially more false positives than the simulation suggests.

7. Reproducibility

Table 3 documents the simulation parameters used in this study. A competent machine learning engineer should be able to replicate the evaluation from these parameters and the methodology described in Section 3. The simulation framework itself is available on request under MIT license; the detector implementations are not distributed — each stage accepts a callable interface (image: np.ndarray) -> Tuple[float, np.ndarray] returning (confidence, bounding_box) and can be satisfied by any standard inference runtime (across common inference runtimes).

ParameterValue usedNotes
Sequence countSequence count500 per difficulty level (1,500 total)Minimum recommended: 200 for stable F1 estimate
Frame resolutionFrame resolution640 × 480 pxAdjust to match target deployment resolution
Target size rangeTarget size range32–128 px (longest side)Drawn uniformly; covers small-to-medium COCO objects
Occlusion fraction (partial)Occlusion fraction (partial)0.25–0.50Uniform draw within range
Motion blur sigma (adverse)Motion blur sigma (adverse)2.0–4.0 pxGaussian kernel applied to target patch only
Confidence threshold calibration splitCalibration split20% of each difficulty level (held out from test)Platt scaling applied per stage; re-fit if detector changes
Latency measurementLatency measurementWall-clock, 10 warm-up frames discardedCPU-only host; GPU deployments will differ significantly
Random seedRandom seed42 (NumPy default_rng)For sequence generation reproducibility
Table 3 — Reproducibility parameters for the simulated evaluation. A practitioner replicating this study should generate synthetic sequences using the parameters below, then substitute their own detectors at the corresponding stage slots. The simulation framework is intentionally detector-agnostic: it accepts a callable that returns (confidence, bbox) given an image patch, and measures wall-clock latency using the host timer. Threshold calibration (Section 2.3) should be re-run on a held-out calibration split before evaluating on the test split.

All simulation code and configuration files are available on request. Detector implementations are not distributed with the simulation framework — practitioners provide their own via the callable interface described in Section 6.

Practitioners replicating this study on their own detection domains should note two differences from the simulation setup that are likely to change the quantitative results materially: (a) real detector confidence scores are typically less well-calibrated than the simulated scores, making the Platt-scaling step more impactful; and (b) real scene variation introduces spatial correlations across frames that the simulation does not model, which can cause the fallback stage activation rate to be temporally clustered (bursts of hard frames followed by periods of easy frames) rather than independently distributed. The latency implications of this clustering — whether the pipeline budget holds on average across a session — should be evaluated on a per-deployment basis.

References

  1. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., & Zagoruyko, S. (2020). End-to-end object detection with transformers. In Proceedings of ECCV 2020, Lecture Notes in Computer Science, vol. 12346. Springer, Cham. doi:10.1007/978-3-030-58452-8_13
  2. Zhao, Y., Lv, W., Xu, S., Wei, J., Wang, G., Dang, Q., Liu, Y., & Chen, J. (2024). DETRs beat YOLOs on real-time object detection. In Proceedings of CVPR 2024, pp. 16965–16974. IEEE. arXiv:2304.08069
  3. Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: unified, real-time object detection. In Proceedings of CVPR 2016, pp. 779–788. IEEE. doi:10.1109/CVPR.2016.91
  4. Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster R-CNN: towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems 28 (NIPS 2015), pp. 91–99. Curran Associates.
  5. Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., & Zitnick, C. L. (2014). Microsoft COCO: common objects in context. In Proceedings of ECCV 2014, Lecture Notes in Computer Science, vol. 8693. Springer, Cham. doi:10.1007/978-3-319-10602-1_48
  6. Viola, P., & Jones, M. J. (2004). Robust real-time face detection. International Journal of Computer Vision, 57(2), 137–154. doi:10.1023/B:VISI.0000013087.49260.fb
  7. Wang, C.-Y., Bochkovskiy, A., & Liao, H.-Y. M. (2023). YOLOv7: trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of CVPR 2023, pp. 7464–7475. IEEE. arXiv:2207.02696
  8. Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.-Y., & Berg, A. C. (2016). SSD: single shot multibox detector. In Proceedings of ECCV 2016, Lecture Notes in Computer Science, vol. 9905. Springer, Cham. doi:10.1007/978-3-319-46448-0_2
  9. Platt, J. (1999). Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. In Advances in Large Margin Classifiers, pp. 61–74. MIT Press.
  10. Niculescu-Mizil, A., & Caruana, R. (2005). Predicting good probabilities with supervised learning. In Proceedings of ICML 2005, pp. 625–632. ACM. doi:10.1145/1102351.1102430
  11. Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., & Adam, H. (2017). MobileNets: efficient convolutional neural networks for mobile vision applications. arXiv:1704.04861
  12. Tan, M., Pang, R., & Le, Q. V. (2020). EfficientDet: scalable and efficient object detection. In Proceedings of CVPR 2020, pp. 10781–10790. IEEE. arXiv:1911.09070

Cite as: TruPath Labs Research (2026). Multi-Stage Fallback in Real-Time Computer Vision: A Methodology Study. TruPath Labs Preprint TPL-2026-005. trupathventures.net/labs/research/multi-stage-fallback-cv