BODEM: Black-box Detection Explainer

• The paper introduces BODEM, which leverages a hierarchical random masking strategy to produce per-detection saliency maps in a strict black-box setting.
• It applies perturbation-based occlusion and quantifies localization changes using IoU, ensuring that masked regions closely reflect detector sensitivities.
• Empirical evaluations on diverse datasets show BODEM achieves lower deletion AUC and higher insertion AUC than flat masking methods like D-RISE.

Searching arXiv for “BODEM” and the cited ids to ground the article. BODEM, short for Black-box Object Detection Explanation by Masking, is a model-agnostic, post-hoc explainer for deep object detectors that estimates a per-detection saliency map indicating which image regions are most important for a particular predicted bounding box. It was introduced for the strict black-box setting in which only the input image and the detector’s final bounding boxes are available, and it was later situated within the perturbation-based branch of explainable object detection alongside methods such as D-RISE, D-CLOSE, and FSOD. Its defining idea is a hierarchical random masking scheme that moves from coarse exploration to fine refinement so that saliency progressively concentrates on causally relevant structures rather than remaining diffuse and noisy (Moradi et al., 2023, Seyedmomeni et al., 2 Sep 2025).

1. Conceptual definition and scope

BODEM addresses a specific interpretability problem in object detection: given one detected box, determine what the detector saw for that box without using gradients, logits, anchors, proposals, or internal feature maps. In the original formulation, the detector is treated as a pure black box that accepts an image and returns bounding boxes; in the later review literature, this same principle is described as using only the input image and the detector’s final outputs while remaining agnostic to detector architecture (Moradi et al., 2023, Seyedmomeni et al., 2 Sep 2025).

Within the taxonomy of explainable artificial intelligence for object detection, BODEM is a perturbation-based occlusion explainer. Its rationale follows sensitivity analysis: systematically perturb the input, observe how the detector’s output changes, and attribute importance to regions whose masking most strongly alters the target detection. The method is post-hoc, requires no training or fine-tuning, does not fit a surrogate model, and does not use a loss function during explanation generation. A plausible implication is that BODEM is especially relevant in deployment environments where only API-level access to a detector is possible.

Because it operates only on final detections, BODEM is applicable to single-stage detectors such as YOLO, SSD, and EfficientDet, and also to two-stage detectors such as Faster R-CNN. Anchors, proposals, and non-maximum suppression are treated opaquely; attribution is based on changes in final post-NMS boxes rather than on detector-specific internals (Seyedmomeni et al., 2 Sep 2025).

2. Formal problem statement and attribution rule

In the original paper, an object detector $f(I)$ maps an input image $I$ of size $W \times H$ to a set of bounding boxes $O = \{O_1, O_2, \ldots, O_N\}$, where each detection $O_n = (x_1, y_1, x_2, y_2)$ is a rectangle in image coordinates. The goal is to produce, for each detection $O_n$, a saliency map $SM_n \in \mathbb{R}^{W \times H}$ that encodes pixel importance for that detection (Moradi et al., 2023).

A mask is applied by element-wise multiplication. If $M_q \in \{0,1\}^{W \times H}$ is a binary mask, the masked image is

$I'_q = I \circ M_q,$

where masked pixels are set to zero. The original paper uses zero as background fill; the review notes that blending with a baseline such as a gray value is also a practical possibility in black-box perturbation pipelines (Moradi et al., 2023, Seyedmomeni et al., 2 Sep 2025).

For a target detection $O_n$ in the original image and a matched detection $I$0 in the masked image, BODEM defines similarity through Intersection over Union:

$I$1

The original implementation then converts this overlap into an importance score for each masked block:

$I$2

This means that a block becomes important when masking it causes the matched box to deviate strongly from the original detection. The later review presents the same principle more generally as black-box attribution by comparing original and perturbed outputs through IoU, either as retained overlap or as overlap lost under occlusion. In both descriptions, attribution is driven by localization change rather than by class-score change (Seyedmomeni et al., 2 Sep 2025).

At the block level, BODEM averages importance over all masks in which a block appears:

$I$3

This blockwise aggregation is the basis for the hierarchical saliency update performed across successive masking levels.

3. Hierarchical random masking algorithm

BODEM proceeds through three repeated stages at each hierarchy level: mask generation, model inquiry, and saliency estimation with refinement. The hierarchy begins with coarse image partitions and iteratively halves block size, moving from broad localization to sharper attribution (Moradi et al., 2023).

At level $I$4, the image is partitioned into non-overlapping square blocks of size $I$5. At level $I$6, block width and height are halved relative to level $I$7. In the experiments reported in the original paper, six levels were used, with block sizes:

• $I$8
• $I$9
• $W \times H$0
• $W \times H$1
• $W \times H$2
• $W \times H$3

Mask generation is not uniform over the whole image at every level. At the coarsest level, all blocks are candidate seeds. At later levels, only blocks with non-zero saliency at the previous level remain candidates. A seed block is selected at random, uniformly at level 1 and with probability proportional to previous-level saliency for $W \times H$4. Its neighbors within distance $W \times H$5 blocks are identified, and 50% of those neighbors are randomly selected to be masked together with the seed. This produces spatially contiguous masked regions rather than isolated random pixels (Moradi et al., 2023).

The saliency update combines prior-level saliency with current-level evidence. Let $W \times H$6 denote the previous map and $W \times H$7 the updated map. With hyperparameters $W \times H$8, the update is

$W \times H$9

In the reported experiments, hyperparameter tuning selected $O = \{O_1, O_2, \ldots, O_N\}$0 and $O = \{O_1, O_2, \ldots, O_N\}$1. The finest-level map is returned as the explanation for the target detection. The original implementation does not apply explicit smoothing beyond assigning block saliency uniformly to all pixels in a block; smoothness emerges from increasingly fine block granularity (Moradi et al., 2023).

The review literature emphasizes the same coarse-to-fine principle in a more general form: levelwise aggregation is normalized, later mask sampling is restricted to a region of interest inferred from prior saliency, and the final attribution is obtained by combining normalized saliency maps across levels. This suggests that the central innovation of BODEM is not masking per se, but hierarchically guided masking that reduces aimless perturbations (Seyedmomeni et al., 2 Sep 2025).

4. Empirical evaluation and reported performance

The original study evaluates BODEM on three object detection settings: a user interface control detection dataset, an airplane detection dataset, and a vehicle detection subset of COCO. The corresponding detectors were YOLO-v5 for UI controls, R-CNN with a VGG-16 backbone for airplanes, and SSD with a ResNet-101 backbone for vehicles (Moradi et al., 2023).

Explanation quality was measured with deletion AUC, insertion AUC, and a convergence score computed from the average pairwise Euclidean distance between saliency maps produced by three runs with different random seeds:

$O = \{O_1, O_2, \ldots, O_N\}$2

Lower deletion AUC is better because salient-pixel removal should rapidly degrade the target detection. Higher insertion AUC is better because re-inserting salient pixels should rapidly restore the detection. Lower convergence is better because it indicates greater stability across random runs (Moradi et al., 2023).

Dataset / detector	BODEM	D-RISE
UI control / YOLO-v5	Deletion 0.058; Insertion 0.875; Convergence 6.051	Deletion 0.113; Insertion 0.612; Convergence 18.406
Airplane / R-CNN	Deletion 0.064; Insertion 0.856; Convergence 7.133	Deletion 0.128; Insertion 0.597; Convergence 19.512
Vehicle / SSD	Deletion 0.069; Insertion 0.860; Convergence 6.420	Deletion 0.137; Insertion 0.605; Convergence 17.381

Qualitatively, the paper reports that BODEM localizes control borders, icons, and relevant text in UI images; head, wings, and tail in airplane detections; and wheels, hood, and boot for cars as well as wheels and seat for motorcycles in vehicle detections, with less extraneous activation than D-RISE. The review later characterizes BODEM’s visual outputs as concentrated, object-specific saliency over structural parts, with fewer spurious highlights outside the target box because of the region-of-interest-focused hierarchy (Moradi et al., 2023, Seyedmomeni et al., 2 Sep 2025).

There is one nomenclatural point worth noting. The abstract of the original BODEM paper states that BODEM outperformed D-RISE and LIME, whereas the detailed results supplied for that paper report quantitative comparisons against D-RISE and explicitly state that LIME is not evaluated in the comparative experiments presented in the details block (Moradi et al., 2023).

5. Relation to other object-detection explainers

The 2025 review places BODEM within a broader landscape of explainability methods for object detection and contrasts it with D-RISE, D-CLOSE, FSOD, and gradient/backpropagation-based methods such as Grad-CAM variants, CRP/LRP, and ODAM (Seyedmomeni et al., 2 Sep 2025).

Relative to D-RISE, BODEM shares the black-box perturbation paradigm but differs in two important respects. First, D-RISE uses many flat random masks sampled across the full image, whereas BODEM restricts later masks to previously salient regions. Second, D-RISE uses a composite weight involving IoU, class-vector similarity, and objectness, while BODEM in its black-box form reduces weighting to IoU-driven localization change. This makes BODEM simpler when only bounding boxes are trusted or accessible, and the review explicitly associates it with higher stability and lower noise than flat random occlusion (Seyedmomeni et al., 2 Sep 2025).

Relative to D-CLOSE, BODEM is simpler because it does not require segmentation into semantic units or superpixels. D-CLOSE uses segmentation at multiple levels, a density-map normalization, and multi-scale map fusion, which the review identifies as useful for varied object sizes and complex textures. A plausible implication is that BODEM is easier to deploy in settings where segmentation quality is uncertain, while D-CLOSE may be preferable when semantically meaningful regions are important (Seyedmomeni et al., 2 Sep 2025).

Relative to FSOD, BODEM remains strictly black-box. FSOD learns an explainer that approximates Shapley attributions using internal feature maps and a query map $O = \{O_1, O_2, \ldots, O_N\}$3; it is fast at test time and the review attributes to it strong energy-based pointing game and insertion performance with real-time speed. BODEM, by contrast, trades speed for stricter operational agnosticism (Seyedmomeni et al., 2 Sep 2025).

Relative to gradient/backpropagation-based methods, BODEM is slower because it requires many forward passes on masked images, but it is also less dependent on architectural access. The review states that gradient and backpropagation methods can be faster and produce sharp maps, yet may lack true causal fidelity under occlusion and require internal access. BODEM is therefore positioned as a method to prefer when strict black-box constraints hold or when insertion/deletion-style fidelity is the primary evaluation target (Seyedmomeni et al., 2 Sep 2025).

6. Robustness, limitations, and practical use

The principal strength attributed to BODEM is robustness through hierarchical refinement. By focusing later mask sampling on high-saliency regions, the method reduces noise from irrelevant occlusions, mitigates sensitivity to mask granularity by adapting cell sizes across levels, and avoids dependence on classifier-score calibration because attribution is based on the target box’s spatial localization rather than on confidence scores (Seyedmomeni et al., 2 Sep 2025).

Its principal limitation is computational cost. If $O = \{O_1, O_2, \ldots, O_N\}$4 is the average detector inference time and $O = \{O_1, O_2, \ldots, O_N\}$5 is the total number of masks, the review gives time complexity as

$O = \{O_1, O_2, \ldots, O_N\}$6

where $O = \{O_1, O_2, \ldots, O_N\}$7 covers IoU matching across detected boxes. Memory usage is dominated by storing masks and the aggregated saliency map, with $O = \{O_1, O_2, \ldots, O_N\}$8 per map plus transient detection outputs. The review judges BODEM to be slower than learned explainers such as FSOD and slower than gradient-based heatmaps, though more efficient in saliency-to-cost ratio than flat random perturbation methods such as D-RISE because the hierarchy reduces aimless masking (Seyedmomeni et al., 2 Sep 2025).

Several failure modes are explicitly identified. Over-occlusion artifacts can appear when coarse masks are too large and destabilize detections. Highly overlapping instances of the same class can produce ambiguous attribution because box matching may conflate neighboring instances when the detector shifts or merges boxes under occlusion. Class imbalance is not directly addressed. In addition, because BODEM relies solely on IoU, it emphasizes spatial fidelity over class semantics and does not produce signed attributions that separate positive from negative evidence (Seyedmomeni et al., 2 Sep 2025).

For reproducibility, the review recommends specifying the mask pyramid $O = \{O_1, O_2, \ldots, O_N\}$9, the occlusion semantics (binary occlusion or soft attenuation), the matching policy for target selection and missing detections, and the aggregation and normalization scheme. It also notes that the paper does not provide code. In practical deployment, BODEM is recommended for validation, debugging, and safety analysis in domains such as autonomous driving, medical imaging, and security systems when gradients or model internals are inaccessible, but its maps should be checked against possible occlusion artifacts before being used in high-stakes decisions (Seyedmomeni et al., 2 Sep 2025).

BODEM thus occupies a specific niche in explainable object detection: it is a high-fidelity, model-agnostic localization explainer for settings where only final detections are exposed. Its contribution lies less in inventing occlusion-based attribution itself than in structuring that attribution hierarchically so that a detector’s response to masking can be converted into cleaner, more stable, and more box-specific saliency maps under genuine black-box constraints (Moradi et al., 2023, Seyedmomeni et al., 2 Sep 2025).