Rethinking Detection Calibration: A Coordinate Perspective¶
Conference: ECCV 2026
Paper: ECCV 2026
Area: Object Detection
Keywords: Confidence calibration, Object detection, Coordinate-wise calibration, Direction estimation, Explainable computer vision
TL;DR¶
Addressing the limitation that conventional detection calibration only measures overall box-level IoU or precision while neglecting independent coordinate alignment and deviation directions, this paper proposes ReDC, a post-hoc calibration framework that integrates coordinate-wise alignment ratio (CAR), confidence re-encoding (CR), and directional displacement estimation (DDE) to deliver fine-grained coordinate-level confidence and directional cues while accurately reconstructing box-level IoU.
Background & Motivation¶
In safety-critical applications such as autonomous driving, medical diagnostics, and robotic surveillance, deep neural network-based object detectors require not only superior localization and categorization accuracy, but also trustworthy confidence estimation. Modern object detectors are notoriously prone to overconfidence, assigning dangerously high confidence scores to predictions that exhibit substantial spatial misalignment or outright classification error. To bridge the gap between predicted confidence scores and true empirical accuracy, confidence calibration has been widely investigated. Unlike image classification where accuracy is binary per category, detection calibration requires defining localization accuracy, which conventional approaches almost universally formulate at the holistic bounding-box level—either as precision under a rigid IoU threshold or by calibrating confidence scores to predict continuous box-level IoU.
However, treating the bounding box as a monolithic calibration unit creates a critical blind spot. Two candidate bounding boxes sharing an identical box-level IoU can exhibit entirely distinct coordinate-level misalignment and spatial error characteristics. For example, a predicted box may align with the ground truth along the vertical borders while drifting drastically on the horizontal borders; alternatively, an upward shift combined with an undersized box makes a detected pedestrian appear further away than reality in autonomous driving, potentially delaying braking actions. Although probabilistic object detectors attempt to model localization uncertainty by predicting Gaussian distributions over bounding box coordinates, they typically collapse coordinate variances back into a single box-level metric, fail to predict the explicit direction of localization errors, and require specialized probabilistic architectures that cannot be easily retrofitted onto mainstream deterministic detectors.
This paper tackles this issue by shifting the calibration paradigm from a holistic box-level view to a fine-grained coordinate and directional perspective. The core idea is to define empirical localization accuracy through a coordinate-wise alignment ratio (CAR), modulate classification logits using bounding box regression features via a confidence re-encoder (CR) to output coordinate-level confidences, estimate deviation directions with a directional displacement estimator (DDE), and bottom-up aggregate these fine-grained scores to accurately approximate box-level IoU.
Method¶
Overall Architecture¶
ReDC is designed as a lightweight, model-agnostic post-hoc calibration framework that seamlessly interfaces with pre-trained, frozen deterministic object detectors. Given an input image, the frozen detector extracts visual features, outputs classification logits, and generates predicted bounding box coordinates. Operating on these frozen representations, ReDC executes two complementary tasks: first, the Confidence Re-encoder (CR) modulates the classification logit with intermediate box regression features to produce independent confidence scores for the four bounding box coordinates; second, the Directional Displacement Estimator (DDE) leverages classification logits and geometric bounding box attributes to predict whether each coordinate overshoots or undershoots the ground truth. Finally, ReDC aggregates coordinate-level confidences and directional information to approximate box-level IoU, achieving both fine-grained coordinate transparency and dependable box-level calibration.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Input Image + Frozen Detector"] --> B["Feature & Prediction Extraction<br/>Classification Logits + Coordinates + Features"]
B --> C["Coordinate-wise Alignment Ratio<br/>Formulate axis intersection & error ratio as target"]
B --> D["Confidence Re-encoding<br/>Box regression features modulate logits into 4 coordinate scores"]
B --> E["Directional Displacement Estimation<br/>Geometric attributes & logits predict coordinate deviation signs"]
C & D --> F["Coordinate-Level Calibration Optimization<br/>NLL loss aligns coordinate confidence with CAR"]
D & E --> G["Bottom-up IoU Approximation<br/>Aggregate fine-grained scores with Isotonic Regression or Platt Scaling"]
Key Designs¶
1. Coordinate-wise Alignment Ratio: Quantifying Per-Coordinate Empirical Accuracy Traditional calibration collapses localization into a single IoU value, obscuring the heterogeneity of localization errors across axes. To establish an empirical accuracy ground truth for each coordinate, the paper introduces the Coordinate-wise Alignment Ratio (CAR). For a predicted box \((\hat{x}_1, \hat{y}_1, \hat{x}_2, \hat{y}_2)\) and ground-truth box \((x_1, y_1, x_2, y_2)\), the coordinate-wise absolute distance is defined as \(\text{dist}_{x^l} = |\hat{x}^l - x^l|\) and \(\text{dist}_{y^l} = |\hat{y}^l - y^l|\) for corner indices \(l \in \{1, 2\}\), while the axis-wise intersection spans are given by: $\(\text{inter}_w = \max(0, \min(\hat{x}_2, x_2) - \max(\hat{x}_1, x_1)), \quad \text{inter}_h = \max(0, \min(\hat{y}_2, y_2) - \max(\hat{y}_1, y_1))\)$ Using these quantities, the CAR values \(\bar{\mathbf{p}} = (\bar{\text{p}}_{x_1}, \bar{\text{p}}_{y_1}, \bar{\text{p}}_{x_2}, \bar{\text{p}}_{y_2})\) are formulated as: $\(\bar{\text{p}}_{x^l} := \frac{\text{inter}_w}{\text{dist}_{x^l} + \text{inter}_w}, \quad \bar{\text{p}}_{y^l} := \frac{\text{inter}_h}{\text{dist}_{y^l} + \text{inter}_h}\)$ Each CAR score naturally lies within \([0, 1]\). When the bounding boxes do not overlap horizontally (\(\text{inter}_w = 0\)), \(\bar{\text{p}}_{x^l} = 0\); conversely, when the predicted coordinate perfectly aligns with the ground truth (\(\text{dist} = 0\)), \(\bar{\text{p}}_{x^l} = 1.0\). Perfect coordinate calibration is thus formally achieved when the predicted coordinate confidence equals the expectation of empirical CAR across all predictions with that confidence.
2. Confidence Re-encoding: Modulating Classification Logits with Localization Features Standard object detectors predict a single classification score that lacks spatial sensitivity across different corners. The Confidence Re-encoder (CR) module \(\phi_{\text{CR}}\) is constructed as a compact multi-layer perceptron. It takes the penultimate bounding box regression feature representation \(\hat{\mathbf{f}}_i = \phi_R[:-2](\phi_F(x_i))\) as input to generate coordinate-specific temperature scalings and learnable biases, re-encoding the classification logit vector \(\hat{\mathbf{z}}_i\): $\(\hat{\mathbf{p}}_{(t, i)} = \sigma\left(\frac{\hat{\mathbf{z}}_i}{\phi_{\text{CR}_t}(\hat{\mathbf{f}}_i)} + \beta_t\right), \quad t \in \{x_1, y_1, x_2, y_2\}\)$ where \(\sigma(\cdot)\) is the sigmoid activation function and \(\beta_t\) represents a coordinate-specific trainable bias. This formulation disentangles the global categorical confidence into four distinct, localized confidence scores that accurately reflect the boundary alignment quality for left, top, right, and bottom borders.
3. Directional Displacement Estimation: Predicting Spatial Deviation Vectors Symmetric uncertainty metrics only indicate magnitude but cannot distinguish whether a predicted border overshoots or undershoots the target object, which is vital for navigation and obstacle avoidance. The Directional Displacement Estimator (DDE) \(\phi_{\text{DDE}}\) resolves this by predicting the deviation direction \(\bar{s}_t \in \{+1, -1\}\) of the ground truth relative to the prediction. DDE uses an MLP that takes as input both the classification logit vector \(\hat{\mathbf{z}}_i\) and geometric attributes of the predicted box, including center coordinates \((\hat{\text{cx}}_i, \hat{\text{cy}}_i)\), width \(\hat{w}_i\), height \(\hat{h}_i\), area \(\hat{A}_i\), and aspect ratio \(\hat{R}_i\). During inference, the continuous predicted logit is converted into a binary direction prediction against a class-specific threshold \(\tau_t^c\): $\(\hat{s}_{(t, i)} = \begin{cases} +1 & \text{if } \sigma(\hat{z}^c_{s_{(t, i)}}) > \tau_t^c \\ -1 & \text{otherwise} \end{cases}\)$ By coupling directional signs with coordinate confidence scores, downstream systems receive actionable spatial guidance regarding where and in which direction a predicted boundary needs adjustment.
4. Bottom-up IoU Approximation: Unifying Coordinate and Box-Level Calibration To retain full backward compatibility with standard box-level decision workflows and evaluation metrics, ReDC demonstrates that box-level IoU can be analytically derived from CAR and directional displacement. When the coordinate-wise expected calibration error (C-ECE) reaches its theoretical minimum, the calibrated confidence scores converge to the true CAR. ReDC directly approximates box IoU using the calibrated coordinate confidences and predicted deviation directions, followed by a light post-hoc fit using Isotonic Regression (IR) or Platt Scaling (PS) on the validation set. This bridges fine-grained coordinate-level transparency with robust box-level calibration.
Loss & Training¶
ReDC follows an efficient two-stage post-hoc training scheme while keeping the primary object detector completely frozen: 1. CR Training: Trained via negative log-likelihood (NLL) loss to align continuous coordinate confidence predictions \(\hat{\mathbf{p}}\) with the empirical ground-truth CAR scores \(\bar{\mathbf{p}}\): $\(\mathcal{L}_{\text{cal}} = \mathbb{E}\left[ - \sum_{t} \left(\bar{\text{p}}_t \log(\hat{\text{p}}_t) + (1 - \bar{\text{p}}_t)\log(1 - \hat{\text{p}}_t)\right) \right]\)$ 2. DDE Training: With the confidence re-encoder fixed, DDE is trained using binary cross-entropy (BCE) loss on the directional targets \(\bar{s}_t \in \{+1, -1\}\). The class-specific decision thresholds \(\tau_t^c\) are determined by optimizing directional accuracy on the validation split. 3. Evaluation Metrics: Introduces Coordinate-wise Expected Calibration Error (C-ECE, Eq. 14), which measures the binned absolute calibration error between predicted coordinate confidence and average CAR, and Direction-aware Calibration Error (Da-CE, Eq. 17), which quantifies the distance between true directional misalignment \(\bar{\mathbf{U}} = \bar{s} \cdot \bar{\text{p}}\) and predicted directional mismatch \(\hat{\mathbf{U}} = \hat{s} \cdot \hat{\text{p}}\) over true positive predictions.
Key Experimental Results¶
Main Results¶
Experiments were conducted on COCO (80 classes) and Cityscapes (8 autonomous driving classes) using Deformable-DETR (ResNet-50) as the base detector. Baselines include uncalibrated models, train-time calibration methods (BPC, Cal-DETR), and state-of-the-art post-hoc calibration methods (IR/PS for LaECE0).
Main comparison on the COCO dataset:
| Method | Type | \(C_{x1}\text{-ECE}\downarrow\) | \(C_{y1}\text{-ECE}\downarrow\) | \(C_{x2}\text{-ECE}\downarrow\) | \(C_{y2}\text{-ECE}\downarrow\) | D-ECE \(\downarrow\) | \(\text{LaECE}_0\downarrow\) | AP \(\uparrow\) | LRP \(\downarrow\) |
|---|---|---|---|---|---|---|---|---|---|
| Uncalibrated (Deformable-DETR) | - | 17.3 | 17.4 | 17.4 | 17.2 | 14.9 | 12.7 | 51.3 | 57.3 |
| TCD (NeurIPS'22) | Train-time | 18.0 | 18.1 | 18.1 | 17.5 | 14.4 | 13.1 | 51.3 | 57.1 |
| BPC (CVPR'23) | Train-time | 15.5 | 15.3 | 15.3 | 15.2 | 11.3 | 12.8 | 50.3 | 58.4 |
| Cal-DETR (NeurIPS'23) | Train-time | 14.0 | 13.9 | 13.9 | 13.6 | 9.8 | 11.7 | 52.5 | 56.2 |
| IR for \(\text{LaECE}_0\) (ECCV'24) | Post-hoc | 12.0 | 11.9 | 11.8 | 11.9 | 2.4 | 7.8 | 51.0 | 57.3 |
| PS for \(\text{LaECE}_0\) (ECCV'24) | Post-hoc | 13.9 | 14.0 | 14.0 | 13.8 | 2.3 | 9.7 | 51.3 | 57.3 |
| ReDC (Ours, IR) | Post-hoc | 7.7 | 10.5 | 11.4 | 10.6 | 2.5 | 7.9 | 50.4 | 57.4 |
| ReDC (Ours, PS) | Post-hoc | 7.6 | 10.4 | 10.9 | 10.9 | 2.4 | 10.0 | 50.3 | 57.3 |
Main comparison on the Cityscapes autonomous driving dataset:
| Method | Type | \(C_{x1}\text{-ECE}\downarrow\) | \(C_{y1}\text{-ECE}\downarrow\) | \(C_{x2}\text{-ECE}\downarrow\) | \(C_{y2}\text{-ECE}\downarrow\) | D-ECE \(\downarrow\) | \(\text{LaECE}_0\downarrow\) | AP \(\uparrow\) | LRP \(\downarrow\) |
|---|---|---|---|---|---|---|---|---|---|
| Uncalibrated | - | 13.7 | 15.5 | 14.2 | 15.5 | 13.7 | 13.4 | 44.5 | 66.4 |
| Cal-DETR | Train-time | 13.4 | 15.1 | 13.4 | 14.5 | 13.0 | 12.7 | 38.7 | 70.6 |
| IR for \(\text{LaECE}_0\) | Post-hoc | 12.2 | 13.6 | 11.6 | 13.9 | 1.5 | 7.5 | 43.5 | 66.4 |
| PS for \(\text{LaECE}_0\) | Post-hoc | 12.9 | 14.7 | 13.1 | 15.1 | 1.0 | 9.6 | 44.5 | 66.4 |
| ReDC (Ours, IR) | Post-hoc | 6.9 | 11.1 | 10.3 | 11.3 | 1.2 | 6.5 | 42.5 | 67.0 |
| ReDC (Ours, PS) | Post-hoc | 6.5 | 11.4 | 10.5 | 10.6 | 1.1 | 9.4 | 42.0 | 66.9 |
Ablation Study¶
The impact of Directional Displacement Estimation (DDE) was ablated on COCO across five random seeds:
| Directional Estimation (DDE) | \(C_{x1}\text{-ECE}\downarrow\) | \(C_{y1}\text{-ECE}\downarrow\) | \(C_{x2}\text{-ECE}\downarrow\) | \(C_{y2}\text{-ECE}\downarrow\) | \(\text{LaECE}_0\downarrow\) | \(\text{LaACE}_0\downarrow\) |
|---|---|---|---|---|---|---|
| Without DDE | 11.22 (±0.058) | 11.90 (±0.106) | 10.42 (±0.054) | 10.62 (±0.134) | 9.96 (±0.022) | 23.14 (±0.034) |
| With DDE (Full ReDC) | 11.18 (±0.038) | 10.86 (±0.042) | 10.26 (±0.074) | 10.58 (±0.106) | 9.88 (±0.022) | 23.14 (±0.002) |
Model-agnostic evaluation across diverse detector architectures on COCO: - VFNet (One-stage): Da-CE decreased from 48.1% to 25.7%, while \(C_{x1}\text{-ECE}\) dropped from 22.9% to 12.9%; - Cascade R-CNN (Two-stage): Da-CE dropped from 42.9% to 26.0%, and \(C_{x1}\text{-ECE}\) improved from 18.3% to 10.6%; - DINO ViT (Transformer-based): Da-CE was reduced from 19.3% to 28.1% baseline error, with \(C_{x1}\text{-ECE}\) reaching 9.4%.
Key Findings¶
- Fine-grained Calibration Superiority: On COCO, ReDC (IR) lowers the \(x_1\) calibration error \(C_{x1}\text{-ECE}\) to 7.7% (compared to 12.0% for IR for \(\text{LaECE}_0\) and 17.3% uncalibrated), while maintaining an average C-ECE of 10.0%, demonstrating superior alignment with individual coordinate geometry.
- Uncompromised Box-Level Quality: ReDC preserves box-level calibration, obtaining 7.9% \(\text{LaECE}_0\) on COCO and outperforming all baselines on Cityscapes with 6.5% \(\text{LaECE}_0\) and 1.2% D-ECE.
- Robust Domain Shift Generalization: Under domain corruption, ReDC maintains an average C-ECE of 14.1% on COCO-C and 12.0% on Foggy Cityscapes, outperforming traditional calibrators that suffer sharp degradation on coordinate metrics. Moreover, analysis reveals that in boxes with top-3% \(y_2\) coordinate error, ReDC's calibrated confidence exhibits an inverse correlation with pixel distance error, accurately reflecting positional reliability.
Highlights & Insights¶
- From Holistic Black-Box to Coordinate Decomposition: Demonstrates that identical box-level IoU conceals severe coordinate-wise error discrepancies, establishing a paradigm where calibration addresses both continuous alignment magnitude and discrete deviation direction.
- Lightweight, Plug-and-Play Architecture: Requires no detector retraining or architectural alterations, introducing two minimal MLPs that operate post-hoc on existing feature maps and logits across one-stage, two-stage, and transformer detectors.
- Novel Benchmark Metrics (C-ECE & Da-CE): Fills a critical methodological void in detection calibration by introducing metrics that explicitly penalize directional blindness and coordinate-level misalignment.
Limitations & Future Work¶
- Non-Rigid Geometry and Severe Occlusion: CAR and DDE rely on axis-aligned bounding box projections, which can misalign with semantic object boundaries under severe deformation or complex occlusions.
- False Positive Handling in Directional Metrics: Because false positives lack matched ground-truth boxes, Da-CE zeroes out direction errors on FPs, evaluating directions only over true positive detections.
- Absence of a Closed-Loop Box Refinement: ReDC operates strictly as an uncertainty assessment module; extending predicted directional displacements and coordinate confidences to iteratively refine bounding box coordinates during inference remains an open and promising direction.
Related Work & Insights¶
- vs Kuzucu et al. (ECCV 2024, LaECE0): Kuzucu et al. aligned confidence to continuous box IoU but retained the box as a single atomic unit. ReDC proves that IoU is analytically derivable from CAR and directional displacement, linking coordinate-level and box-level calibration.
- vs GP-Normal / Küppers et al. (ECCVW 2022): Probabilistic approaches estimate coordinate variance but fail to predict signed directional bias and require specialized probabilistic detector designs. ReDC applies to standard deterministic detectors and achieves significantly higher error correlation (0.1771 vs 0.0263).
- vs Cal-DETR / BPC (NeurIPS 2023 / CVPR 2023): These train-time calibration frameworks introduce custom loss functions during detector pretraining, which is computationally prohibitive and model-specific. ReDC provides superior coordinate calibration post-hoc with zero training overhead on the primary detector.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Pioneering coordinate-level and direction-aware calibration for deterministic object detectors.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Thorough validation across COCO, Cityscapes, corrupted domains, multiple detector families, and detailed ablation/correlation studies.
- Writing Quality: ⭐⭐⭐⭐⭐ Rigorous mathematical formulation, clear prose, and compelling visual evidence.
- Value: ⭐⭐⭐⭐⭐ Highly practical, plug-and-play post-hoc solution with immediate safety implications for robotics and autonomous driving.