ReGT: Vision Reannotation & Temporal Forecasting

• ReGT is a dual-use term, representing both refined ImageNet reannotation for realistic evaluation and a regional temporal graph model for truck parking forecasting.
• The reannotation process in ReGT reduces label noise, enables multilabel evaluation, and improves top-1 accuracy by up to 10.8 percentage points.
• The RegT-GCN component segments geographic graphs into regional subgraphs, achieving 10–20% lower forecasting errors compared to traditional whole-graph approaches.

ReGT is a term that appears in two distinct research contexts: (1) as "Re-annotated Ground Truth" in large-scale computer vision benchmarks, and (2) as the shorthand for the Regional Temporal Graph Convolutional Network (RegT-GCN) in spatio-temporal forecasting on geographic graphs. Each usage is associated with substantial methodological advances in their respective domains.

1. Definition and Scope

1.1. ReGT: Re-annotated Ground Truth

ReGT ("Re-annotated Ground Truth") is a comprehensive, human-curated, multilabel reannotation of 625 ImageNet-1k classes applied to the 50,000-image validation subset. The primary goal is to correct label errors, admit multiple valid labels per image, and provide a more realistic (“real-world”) evaluation target for vision and vision-LLMs. In contrast to the original single-label ImageNet ground truth (ImGT), ReGT accommodates 0, 1, or several admissible labels per image and addresses ambiguity with equivalence classes for synonyms and near-duplicates (Kisel et al., 6 Mar 2026).

1.2. ReGT in Spatio-Temporal Graphs: Regional Temporal GCN

In the graph neural network literature, ReGT denotes the Regional Temporal Graph Convolutional Network (RegT-GCN), a model developed for multi-site truck parking usage prediction. ReGT partitions large spatial graphs into regional subgraphs (often at the state level), applies dedicated GCNs per region, fuses the region-specific embeddings through GCN-augmented recurrent units, and predicts future occupancy rates for each site (Tamaru et al., 2024).

2. Motivations and Rationale

2.1. ReGT (Re-annotated Ground Truth)

Several studies observed inconsistencies in multimodal LLM (MLLM) benchmarking due to evaluation artifacts—discrepant protocols, noisy single-label ground truth, and misclassification of plausible predictions as errors. ReGT was introduced to eliminate annotation artifacts by:

• Allowing true multilabeling required for complex visual scenes.
• Systematically correcting label omissions and merging ambiguous or synonymous categories.
• Reducing spurious penalization of models due to limitations in single-label protocols (Kisel et al., 6 Mar 2026).

2.2. ReGT (Regional Temporal GCN)

Accurate profiling of truck parking usage requires models to capture both spatial dependencies (adjacent parking sites, shared corridors) and temporal dynamics (diurnal/weekly usage fluctuations). Regional decomposition addresses the over-smoothing and diluted regional patterns inherent in large, heterogeneously connected graphs:

• Smaller, intra-region subgraphs strengthen local information propagation.
• Regionally defined decompositions yield sharper forecasts compared to random or entire-graph GCNs.
• Reduced computational complexity due to lower node degrees within subgraphs (Tamaru et al., 2024).

3. Methodological Frameworks

3.1. ReGT: Annotation Process and Data Structure

• Coverage: 31,250 images (625 classes × 50 images/class).
• Exclusions: 375 expert-obscure classes unchanged; specialist reannotation (e.g., Mustelidae) used where feasible.
• Annotation Protocol:
  • Annotators were trained on challenging ImageNet cases and equipped with custom tools presenting all class names and top-20 vision model predictions.
  • For each image, annotators selected all present categories drawn from the 1,000 ImageNet classes, without restriction on label count.
  • Ambiguity resolved by equivalence sets for synonyms and near-duplicates; images lacking any recognized category are assigned |L_i|=0.
  • Difficult or disputed cases underwent adjudication by expert annotators with blinded access to human, ImGT, and MLLM outputs (Kisel et al., 6 Mar 2026).

3.2. ReGT (RegT-GCN): Model Architecture

• Input Graph Construction:
  • Nodes: Truck parking sites, with features including static (capacity, amenities), dynamic (occupancy rate), and temporal (time-of-week) attributes.
  • Edges: Sites within 40 miles connected, weighted by actual driving distance.
• Regional Decomposition:
  • The graph $G$ is split into $R$ regionally defined subgraphs $M^r$, each representing a geographic cohort (e.g., a US state).
• Forward Pass:
  • Extract spatial features with a GCN over $G$, decompose to $M^r$, apply a region-specific GCN, concatenate region-wise outputs.
  • 2. The sequence of embeddings over the history window is fused using an attention-weighted GCN-GRU.
  • 3. A decoder transforms the resulting hidden state(s) to predict short- and medium-term (10 to 360 min) occupancy rates per site (Tamaru et al., 2024).

4. Statistical Outcomes and Evaluations

4.1. Annotation Quality and Impact (ReGT as Ground Truth)

Key improvements realized by ReGT include:

• Label Noise Reduction: ImGT single-label omission rates fell from ~20% (Northcutt et al. 2021) to 9.6% within the reannotated subset.
• Label Structure: Of 31,250 images, 57.8% retained single labels, 37.9% received multilabels, and 4.3% were assigned zero labels.
• Inter-annotator Adjudication: Difficult cases cross-validated by experts, achieving ~50% error correction when integrating MLLM suggestions (e.g., GPT-4o).
• Evaluation Protocol Impact: Top-1 accuracy for leading MLLMs improved by 6–10.8 pp, with models most sensitive to annotation quality seeing the largest gains. Vision-only baselines received smaller improvements (<4.1 pp), illustrating the relevance of label noise primarily for less supervised paradigms (Kisel et al., 6 Mar 2026).

4.2. Forecasting Accuracy (RegT-GCN)

The regional decomposition strategy produced substantial improvements over baselines:

• Metrics (30 min horizon): RMSE 0.094 (vs. 0.115 T-GCN), MAE 0.067 (vs. 0.089), MAPE 9.36% (vs. 12.78%).
• Across all horizons: Error rates were 10–20% lower than alternatives, outperforming ablations without regionalization (random or single-graph GCNs).
• Ablation Study: Random subgraph division (RanT-GCN) outperformed whole-graph baselines but underperformed regional (state-wise) splits, confirming the importance of geo-aligned decomposition (Tamaru et al., 2024).

Model/Protocol	RMSE (30 min)	MAE (30 min)	MAPE (30 min)
RegT-GCN (ReGT)	0.094	0.067	9.36%
T-GCN	0.115	0.089	12.78%
RanT-GCN	(not given)	(not given)	(improved, but less than RegT-GCN)

5. Practical Applications and Insights

5.1. ReGT for Benchmarking and Curation

• Model Comparison: ReGT narrows the gap between MLLM and highly tuned supervised vision models, demonstrating that much reported underperformance is attributable to flawed labels and protocol rather than intrinsic model limits.
• Task Sensitivity: Evaluations such as Open-World (OW), Closed-World (CW+), and Multiple-Choice (MC) demonstrate that Open-World protocol’s perceived underperformance is partly eliminated by more accurate ground truth.
• Human–MLLM Curation: In expert annotation, MLLM outputs (e.g., GPT-4o) were integrated or confirmed in >50% of difficult cases, establishing their utility for large-scale reannotation workflows (Kisel et al., 6 Mar 2026).

5.2. ReGT for Spatio-Temporal Decision Support

• Infrastructure Management: RegT-GCN's predictive accuracy facilitates more effective truck park management, allocation of new sites, and anticipatory regulation (e.g., under HOS constraints).
• Scalability: Regionalized modeling decreases computational costs, supports multi-state deployments, and preserves region-specific behavioral and temporal signatures (Tamaru et al., 2024).

6. Illustrative Examples and Case Studies

6.1. Fine-Grained Visual Annotation

• Equivalence Classes: Ambiguity is codified through equivalence sets: "laptop computer" vs "notebook computer", "sunglasses" vs "sunglass", ensuring that different synonymous labels do not yield penalization.
• Annotation Divergence: Scenes previously assigned a single label ("camera") are now multilabel ("camera", "lens"); polysemous or ambiguous visuals are captured with more fidelity.
• Specialist Subsets: For obscure classes (e.g., Mustelidae), expert reannotation increases recall from 66%–100% (original) to 92%–100% (expert), particularly benefitting MLLMs and VLMs (Kisel et al., 6 Mar 2026).

6.2. Spatio-Temporal Model Workflows

Algorithmic example for RegT-GCN:

for t in prediction_horizon: for k in range(K): H_t_minus_k = StructuralGCN(G_{t-k}, X_{t-k}) regional_embeddings = [] for region in regions: m_r = RegionalGCN(M^r_{t-k}, X^r_{t-k}) regional_embeddings.append(m_r) gamma_t_minus_k = concatenate(regional_embeddings) temporal_embedding = GCN_GRU([gamma_{t-k} for k in range(K)]) X_hat_{t+1} = Decoder(temporal_embedding)

(Tamaru et al., 2024)

A plausible implication is that such modular workflows—regional decomposition, spatial feature extraction, temporal fusion—can generalize to other infrastructure-level demand forecasting domains.

7. Limitations and Future Directions

• Annotation Cost: The high resource commitment for comprehensive human-driven multilabel annotation restricts the scale and frequency of updates to resources such as ReGT.
• Regional Decomposition Choice: Defining optimal regional boundaries is non-trivial; while state-level splits perform well, urban vs. rural gradients or traffic-driven partitions may offer further gains.
• Generalization: Label structure and regional graph design are problem-specific; transfer to other vision or spatio-temporal tasks may require task-tailored modifications (Kisel et al., 6 Mar 2026, Tamaru et al., 2024).

ReGT, in both contexts, demonstrates that principled redesign of ground truth or graph structure yields substantial advances in model reliability, benchmarking, and real-world utility.