MONO Framework: Unified Approaches

• MONO framework is a set of advanced, domain-specific methodologies that enable real-time monocular depth estimation, robust vulnerability dataset construction, and sensitive collider searches.
• It leverages lightweight architectures, self-supervised learning, and geometric reasoning to achieve state-of-the-art performance with minimal computational overhead.
• Empirical results show notable improvements in metrics such as Abs Rel, AP3D, and vulnerability detection accuracy, underscoring its practical real-world deployability.

The term "MONO framework" encompasses several distinct, technically mature frameworks spanning computer vision, program analysis, and collider phenomenology. In each field, “MONO” or “Mono-” refers to a specific methodological innovation—either in model design, dataset construction, or experimental signal extraction. The following sections comprehensively detail these frameworks, focusing on their architectures, methodologies, empirical results, and research significance, as presented in "RTS-Mono: A Real-Time Self-Supervised Monocular Depth Estimation Method for Real-World Deployment" (Cheng et al., 18 Nov 2025), "Mono: Is Your 'Clean' Vulnerability Dataset Really Solvable?" (Gao et al., 4 Jun 2025), "Mono-ViFI: A Unified Learning Framework for Self-supervised Single- and Multi-frame Monocular Depth Estimation" (Liu et al., 2024), "EFT analysis of leptophilic dark matter at future electron-positron colliders in the mono-photon and mono-$Z$ channels" (Kundu et al., 2021), and "MonoGRNet: A General Framework for Monocular 3D Object Detection" (Qin et al., 2021).

1. Monocular Depth Estimation Frameworks

RTS-Mono: Real-Time Self-Supervised Monocular Depth

RTS-Mono is a lightweight, fully convolutional encoder–decoder architecture targeting real-time, self-supervised monocular depth estimation for deployment contexts demanding high speed and low model size. The encoder, derived from Lite-Encoder and integrated with hybrid CNN+Transformer blocks, processes an input RGB image $I \in \mathbb{R}^{H \times W \times 3}$ and yields four hierarchical feature maps $(F_0, F_1, F_2, F_3)$ with channel dimensions $(48, 48, 80, 128)$. The decoder sparsely fuses only the deepest features $(F_1, F_2, F_3)$ via a cascade of convolution, lightweight fusion, and attention-based upsampling blocks, minimizing redundancy and computational overhead. The final inverse depth map, $D^{-1}$, is synthesized by upsampling and concatenation with the shallowest feature $F_0$.

Training is self-supervised, utilizing photometric reprojection between temporally adjacent frames, an edge-aware smoothness regularizer, and a cross-scale depth consistency term. The composite loss function is

$$L = \gamma L_p + \beta L_s + \lambda L_d,\quad (\gamma=1.0,\,\beta=10^{-3},\,\lambda=1.0).$$

RTS-Mono achieves state-of-the-art performance among sub-4M parameter models, with Abs Rel and Sq Rel outperforming prior lightweight methods by 5.6% and 9.8% at low resolution and 6.1% and 1.9% at high resolution. Real-world validation includes 49 FPS inference speed on NVIDIA Jetson Orin with robust generalization in dynamic driving scenes (Cheng et al., 18 Nov 2025).

Mono-ViFI: Unified Single- and Multi-frame Self-supervised Depth

Mono-ViFI introduces a unified self-supervised architecture capable of both single-frame and multi-frame monocular depth estimation. The backbone encoder (e.g., ResNet-18, D-HRNet) is shared, while decoders are distinct for single and multi-frame predictions. Temporal augmentation via a frozen flow-based video frame interpolation (VFI) module creates intermediate "virtual" camera views. This enables the VFI-assisted multi-frame fusion module to reliably align and fuse features from temporally separated frames using motion and occlusion information.

The training objective combines photometric reconstruction loss, edge-aware smoothness, spatial affine data augmentation, and a novel triplet regularization enforcing depth consistency among single-frame, multi-frame, and augmented predictions: $$L(t) = L_{ss}(t) + L^m_{ss}(t) + \tilde{L}_{ss}(t) + \lambda L_{tc}(t),\qquad \lambda = 0.2.$$

Mono-ViFI demonstrates substantial improvements over single-view and many-view baselines with only minimal computational overhead, achieving Abs Rel values as low as 0.091 (multi-frame, D-HRNet, KITTI) (Liu et al., 2024).

2. Frameworks for Automated Program Vulnerability Dataset Construction

mono: LLM-powered Construction for Solvable Vulnerability Datasets

The mono framework is a multi-agent, LLM-driven system for constructing high-integrity vulnerability datasets by simulating human expert reasoning. It targets issues such as inaccurate patch labeling, insufficient contextual information, and the presence of "undecidable" patches in datasets like MegaVul.

The architecture comprises the following pipeline:

1. Semantic-aware Patch Classification: An LLM classifier processes code diffs, commit messages, and pull request discussions to assign a “SecurityFix” label with a calibrated confidence threshold $\tau$ (e.g., 0.9). Pseudocode: $(F_0, F_1, F_2, F_3)$8
2. Iterative Contextual Analysis: An LLM "AnalysisAgent" iteratively identifies missing program context ("GAPs"); a "ContextAgent" then fetches code facts (e.g., definitions, call graphs, data flows) via Joern CPG queries. This loop continues until a confident root cause explanation is produced or $I \in \mathbb{R}^{H \times W \times 3}$0 iterations are reached.
3. Systematic Root-cause Validation and Undecidable Patch Detection: If no valid trigger chain can be established, the sample is flagged as undecidable.

Empirical evaluation on MegaVul indicates that mono corrects 31% of labeling errors, recovers 89% of inter-procedural vulnerabilities, and refines the dataset by excluding approximately 16.7% undecidable cases. This results in an average 15% gain in downstream model accuracy for vulnerability detection, particularly when models are given enriched context (Gao et al., 4 Jun 2025).

3. Frameworks for Collider Phenomenology: Mono-photon and Mono-$I \in \mathbb{R}^{H \times W \times 3}$1 Signatures

In collider physics, "MONO" designates searches for physics beyond the Standard Model using final states with missing energy plus a single visible object (photon or $I \in \mathbb{R}^{H \times W \times 3}$2). These signatures are sensitive to invisible particles such as weakly-interacting dark matter.

The Effective Field Theory (EFT) approach for leptophilic dark matter at $I \in \mathbb{R}^{H \times W \times 3}$3 colliders parametrizes all DM-lepton interactions via dimension-6 operators: $I \in \mathbb{R}^{H \times W \times 3}$4 where $I \in \mathbb{R}^{H \times W \times 3}$5 denotes scalar-pseudoscalar (SP), vector-axial (VA), and tensor-axial-tensor (TAT) structures.

Sensitivity projections for the International Linear Collider (ILC, $I \in \mathbb{R}^{H \times W \times 3}$6 TeV, $I \in \mathbb{R}^{H \times W \times 3}$7 ab$I \in \mathbb{R}^{H \times W \times 3}$8) show that mono-photon and mono-$I \in \mathbb{R}^{H \times W \times 3}$9 searches can exclude cutoff scales $(F_0, F_1, F_2, F_3)$0 up to $(F_0, F_1, F_2, F_3)$1 TeV (SP), $(F_0, F_1, F_2, F_3)$2 TeV (VA), and $(F_0, F_1, F_2, F_3)$3 TeV (TAT) at $(F_0, F_1, F_2, F_3)$4 significance for light dark matter. These limits dramatically exceed those accessible through direct/indirect searches and are thus considered the most sensitive probes for leptophilic DM in the low-mass regime (Kundu et al., 2021).

4. Monocular 3D Object Detection

MonoGRNet: Geometric Reasoning for Monocular 3D Detection

MonoGRNet decomposes single-image 3D object detection into four parallel sub-tasks:

1. 2D Detection: Localization and classification of objects in the image plane.
2. Instance-level Depth Estimation (IDE): Regression of the camera-center distance to the object's 3D center.
3. Projected 3D Center Estimation: Regression of the 2D image location of the 3D center.
4. Local Corner Regression: Prediction of the eight 3D bounding box corners in a local object frame.

Each task is addressed via a dedicated network head. The outputs are fused by geometric back-projection to produce full 3D bounding boxes. Weakly supervised learning is enabled by generating pseudo ground-truths for depth and corners using 2D bounding boxes, object priors, and a teacher-student orientation pipeline.

Quantitatively, MonoGRNet achieves $(F_0, F_1, F_2, F_3)$5 (IoU=0.5) of 43.7 (Easy), 36.2 (Mod), and 30.2 (Hard) on KITTI, approaching LiDAR-supervised baselines but within a monocular, proposal-free, real-time architecture ($(F_0, F_1, F_2, F_3)$660 ms/frame) (Qin et al., 2021).

5. Empirical Performance and Benchmarking

The following table summarizes key empirical results for recent "Mono" frameworks:

Framework	Task/Field	Best Metric(s) (KITTI)	Model Size / Speed
RTS-Mono (Cheng et al., 18 Nov 2025)	Monocular Depth Est.	Abs Rel 0.101, δ<1.25 0.897	3M params, 49 FPS (Orin)
Mono-ViFI (Liu et al., 2024)	Unified Depth, Single+Multi	Abs Rel 0.091 (multi, D-HRNet)	1 encoder, 2 decoders, ~real-time
mono (Gao et al., 4 Jun 2025)	Vulnerability Dataset Synth.	+15% F1 gain in vuln. detection (CWE)	~5,573 CVEs, runtime N/A
MonoGRNet (Qin et al., 2021)	Monocular 3D Detection	AP3D 43.7 / 36.2 / 30.2 (Easy–Hard)	0.06s/frame

All results reference dataset-specific protocols (Eigen split for KITTI, MegaVul for security).

6. Impact, Limitations, and Future Directions

These frameworks have advanced the state of the art by integrating architectural efficiency (RTS-Mono, Mono-ViFI), rigorous semantic and contextual reasoning (mono for program repair datasets), or proposal-free, geometry-driven 3D understanding (MonoGRNet). Each provides not only higher accuracy but also practical real-world deployability, either by minimizing compute requirements, improving data quality, or sharpening phenomenological reach.

However, all face domain-specific limitations:

• RTS-Mono, Mono-ViFI, and MonoGRNet remain constrained by monocular geometric ambiguity and non-rigidity in dynamic scenes.
• The mono dataset construction method is limited by static code property graph tooling and cannot handle certain runtime or convention-based vulnerabilities; only ≈12% of ad-hoc LLM queries outside defined wrappers succeed.
• Collider-based MONO-EFT searches are ultimately restricted by the validity domain of the EFT at high momentum transfer.

A plausible implication is that integrating additional modalities (e.g., LiDAR, stereo, runtime monitoring), more expansive tool integration for program reasoning, or full UV-complete model interpretations in collider analyses could further ameliorate failure cases and extend real-world impact. In program analysis, the inclusion of dynamic traces and more advanced semantic querying would address unresolved "undecidable" class patches (Gao et al., 4 Jun 2025). In depth estimation, explicit handling of temporally correlated non-rigid motion may be necessary for optimal generalization (Cheng et al., 18 Nov 2025, Liu et al., 2024).

7. Notational Conventions and Terminological Clarification

"Mono" is contextually overloaded:

• In computer vision, it denotes monocular input, i.e., a single camera or frame; “MonoGRNet”, “RTS-Mono” and “Mono-ViFI” operate on such settings.
• In program analysis, “mono” refers specifically to the dataset refinement and labeling system; the dataset output is denoted "MonoLens."
• In collider physics, "mono-X" (for $(F_0, F_1, F_2, F_3)$7) refers to single visible object plus missing energy searches.

All referenced frameworks are openly available as source code, with links provided in their respective primary publications. For details on formulas, loss definitions, and implementation, see the original arXiv works: (Cheng et al., 18 Nov 2025, Gao et al., 4 Jun 2025, Liu et al., 2024, Kundu et al., 2021, Qin et al., 2021).