Model-Level Mixtures of Object Detectors

• The paper demonstrates that integrating multiple detection components within a unified architecture improves specialization and calibration through expert mixtures.
• It details various methods such as internal probabilistic mixtures, sparse MoE layers, and full-detector gating to combine outputs effectively.
• Balancing losses and routing strategies are key to preventing redundancy and achieving higher AP scores across diverse domains and scales.

Searching arXiv for the cited papers to ground the article in the current record. Model-level mixtures of object detectors are detection systems in which multiple detection components, experts, or full detectors are combined within a learned architecture, rather than only merged after independent inference. In this literature, the mixture can reside in a probabilistic output model over boxes and classes, in a sparse Mixture-of-Experts (MoE) layer embedded inside a detector, or in a gate that weights complete detector branches before decoding and suppression. The common objective is to exploit complementarity while preserving end-to-end learning, adaptive specialization, or explicit expert attribution; conversely, post-hoc reranking and score fusion methods are usually treated as adjacent but distinct paradigms (Yoo et al., 2022, Jain et al., 2023, Varamesh et al., 2019, Pavlitska et al., 20 Apr 2026).

1. Conceptual scope and typology

The literature uses “mixture” at several architectural levels. In one line of work, the mixture is internal to a single detector head: proposals or dense regression outputs are interpreted as mixture components, and training proceeds by likelihood maximization rather than by hard assignment. In another line, sparse MoE layers are inserted into a detector backbone or decoder so that intermediate representations, rather than only final predictions, become expert-specific. A third line treats whole detectors as experts and learns a gate that weights their raw outputs before decoding and suppression. These are all model-level in the sense that the mixture is part of the detector’s computational graph rather than a purely external aggregation rule (Yoo et al., 2022, Varamesh et al., 2019, Jain et al., 2023, Meiraz et al., 17 Nov 2025, Pavlitska et al., 20 Apr 2026).

Mixture locus	Representative formulation	Representative paper
Internal output density	Mixture density over boxes or box parameters	D-RMM (Yoo et al., 2022); Mixture Dense Regression (Varamesh et al., 2019)
Intermediate representation	Sparse MoE in transformer decoder with learned routing	DAMEX (Jain et al., 2023)
Full detector experts	Multiple YOLO detectors with learned gating over raw outputs	YOLOv9-T MoE (Meiraz et al., 17 Nov 2025); domain-specialized YOLO MoE (Pavlitska et al., 20 Apr 2026)

A central boundary in this taxonomy is the distinction between model-level mixtures and late fusion. Detect2Rank learns to rerank detections from DPM, CN, and EES using contextual features and then applies NMS; DBF calibrates black-box detector scores and fuses them with Dempster’s rule; MoCaE calibrates already-trained detectors and refines aggregation with calibrated postprocessing. These methods exploit detector complementarity, but they do not generally route inputs through internal experts in a single jointly structured detector in the same way as D-RMM, DAMEX, or detector-gated YOLO systems (Karaoglu et al., 2014, Lee et al., 2022, Oksuz et al., 2023).

2. Internal probabilistic mixtures within a detector

A direct model-level formulation appears in "End-to-End Multi-Object Detection with a Regularized Mixture Model" (Yoo et al., 2022). For an image $X$ and ground-truth set $g=\{g_1,\dots,g_N\}$, D-RMM treats end-to-end multi-object detection as conditional density estimation: $$p(g_i \mid X) = \sum_{k=1}^{K} \pi_k \,\mathcal{F}(b_i; \mu_k, \gamma_k)\,\mathcal{P}(c_i; p_k).$$ Here, each proposal or query is a mixture component, $\pi_k$ is the mixture coefficient, $\mathcal{F}$ is a box density, and $\mathcal{P}$ is a categorical class distribution. The box coordinates are modeled with a 4D Cauchy distribution with independent coordinates: $$\mathcal{F}(b_i; \mu_k, \gamma_k) = \prod_{j \in D} \mathcal{F}(b_{i,j}; \mu_{k,j}, \gamma_{k,j}), \quad D=\{l,t,r,b\}.$$ The paper states that Cauchy is used because it has heavier tails than Gaussian and is numerically stable, reducing underflow issues.

D-RMM is built on Sparse R-CNN and also applied to AdaMixer. A backbone extracts features, there are $K$ learnable proposal boxes, RoIAlign extracts proposal features, and a D-RMM head predicts

$M^s = \{\pi^s, \mu^s, \gamma^s, p^s, o^s\}$

at stage $s$. At stage $g=\{g_1,\dots,g_N\}$0, the previous stage’s predicted locations $g=\{g_1,\dots,g_N\}$1 are used as new proposals. The scale parameter $g=\{g_1,\dots,g_N\}$2 is made positive via softplus, $g=\{g_1,\dots,g_N\}$3 uses softmax, $g=\{g_1,\dots,g_N\}$4 uses sigmoid, and the mixture coefficient is derived from objectness: $g=\{g_1,\dots,g_N\}$5 At inference, the final-stage $g=\{g_1,\dots,g_N\}$6 is used as box coordinates, the confidence for class $g=\{g_1,\dots,g_N\}$7 is $g=\{g_1,\dots,g_N\}$8, the top-100 predictions are retained, and no NMS is used.

Its training objective is

$g=\{g_1,\dots,g_N\}$9

with a negative log-likelihood term

$$p(g_i \mid X) = \sum_{k=1}^{K} \pi_k \,\mathcal{F}(b_i; \mu_k, \gamma_k)\,\mathcal{P}(c_i; p_k).$$0

and the maximum component maximization loss

$$p(g_i \mid X) = \sum_{k=1}^{K} \pi_k \,\mathcal{F}(b_i; \mu_k, \gamma_k)\,\mathcal{P}(c_i; p_k).$$1

The paper emphasizes that NLL alone allows multiple components to explain the same object, whereas MCM encourages one dominant component per ground truth and suppresses redundant nearby components. The paper further argues that duplicate predictions hurt calibration because confidence becomes dispersed across multiple boxes; with MCM, confidence is concentrated into a single box and Expected Calibration Error improves. On COCO 2017, D-RMM improves Sparse R-CNN from $$p(g_i \mid X) = \sum_{k=1}^{K} \pi_k \,\mathcal{F}(b_i; \mu_k, \gamma_k)\,\mathcal{P}(c_i; p_k).$$2 to $$p(g_i \mid X) = \sum_{k=1}^{K} \pi_k \,\mathcal{F}(b_i; \mu_k, \gamma_k)\,\mathcal{P}(c_i; p_k).$$3 AP with R50, $$p(g_i \mid X) = \sum_{k=1}^{K} \pi_k \,\mathcal{F}(b_i; \mu_k, \gamma_k)\,\mathcal{P}(c_i; p_k).$$4 to $$p(g_i \mid X) = \sum_{k=1}^{K} \pi_k \,\mathcal{F}(b_i; \mu_k, \gamma_k)\,\mathcal{P}(c_i; p_k).$$5 AP with R101, and $$p(g_i \mid X) = \sum_{k=1}^{K} \pi_k \,\mathcal{F}(b_i; \mu_k, \gamma_k)\,\mathcal{P}(c_i; p_k).$$6 to $$p(g_i \mid X) = \sum_{k=1}^{K} \pi_k \,\mathcal{F}(b_i; \mu_k, \gamma_k)\,\mathcal{P}(c_i; p_k).$$7 AP with Swin-T; it improves AdaMixer from $$p(g_i \mid X) = \sum_{k=1}^{K} \pi_k \,\mathcal{F}(b_i; \mu_k, \gamma_k)\,\mathcal{P}(c_i; p_k).$$8 to $$p(g_i \mid X) = \sum_{k=1}^{K} \pi_k \,\mathcal{F}(b_i; \mu_k, \gamma_k)\,\mathcal{P}(c_i; p_k).$$9 AP with R50, $\pi_k$0 to $\pi_k$1 AP with R101, and $\pi_k$2 to $\pi_k$3 AP with Swin-T. On COCO test-dev, the reported gains are about $\pi_k$4 to $\pi_k$5 AP for Sparse R-CNN and $\pi_k$6 to $\pi_k$7 AP for AdaMixer.

A related but more localized formulation appears in "Mixture Dense Regression for Object Detection and Human Pose Estimation" (Varamesh et al., 2019). This work replaces the usual single regression head with a mixture density network. For input $\pi_k$8 and target $\pi_k$9,

$\mathcal{F}$0

with Gaussian component density

$\mathcal{F}$1

For dense object detection, the regression target is $\mathcal{F}$2, and the mixture is applied at each output cell of a CenterNet-style detector. The regression loss is the mixture NLL,

$\mathcal{F}$3

which replaces the standard $\mathcal{F}$4 box regression loss. The total loss is

$\mathcal{F}$5

with $\mathcal{F}$6, $\mathcal{F}$7, and $\mathcal{F}$8.

The main empirical finding is that the mixture components specialize by object scale rather than by category. The Pearson correlation between maximum component choice and object scale is $\mathcal{F}$9, whereas the correlation between components and categories is $\mathcal{P}$0. On COCO validation with the HG backbone, the baseline single-mode model reaches $\mathcal{P}$1 AP, $\mathcal{P}$2 reaches $\mathcal{P}$3 AP, $\mathcal{P}$4 reaches $\mathcal{P}$5 AP, and $\mathcal{P}$6 reaches $\mathcal{P}$7 AP; on COCO test-dev, single-mode HG reaches $\mathcal{P}$8 AP and $\mathcal{P}$9 reaches $$\mathcal{F}(b_i; \mu_k, \gamma_k) = \prod_{j \in D} \mathcal{F}(b_{i,j}; \mu_{k,j}, \gamma_{k,j}), \quad D=\{l,t,r,b\}.$$0 AP; on DLA34, single-mode reaches $$\mathcal{F}(b_i; \mu_k, \gamma_k) = \prod_{j \in D} \mathcal{F}(b_{i,j}; \mu_{k,j}, \gamma_{k,j}), \quad D=\{l,t,r,b\}.$$1 AP and $$\mathcal{F}(b_i; \mu_k, \gamma_k) = \prod_{j \in D} \mathcal{F}(b_{i,j}; \mu_{k,j}, \gamma_{k,j}), \quad D=\{l,t,r,b\}.$$2 reaches $$\mathcal{F}(b_i; \mu_k, \gamma_k) = \prod_{j \in D} \mathcal{F}(b_{i,j}; \mu_{k,j}, \gamma_{k,j}), \quad D=\{l,t,r,b\}.$$3 AP. The paper explicitly states that it does not face mode collapse on COCO, and it argues that the mixture can practically alleviate the need for multi-scale testing. Taken together, these two papers show that model-level mixtures can be instantiated either over the set of object hypotheses itself or over a dense regression subproblem inside the detector. This suggests that “mixture” in detection is not tied to a single architectural granularity.

3. Dataset-aware routing in universal detection

"Dataset-aware Mixture-of-Experts for visual understanding of mixture-of-datasets" (Jain et al., 2023) frames model-level mixtures as a universal detection problem. Its premise is that a universal detector trained on a mixture of datasets should not force all supervision through one shared feed-forward pathway when the datasets differ in domain, scale, label set, and sample count. DAMEX therefore inserts sparse MoE layers into a single DINO detector so that experts specialize at the representation level while the rest of the network remains shared.

The model is built on DINO with an ImageNet-pretrained ResNet-50 backbone and a transformer encoder-decoder. The paper uses a 6-layer encoder and 6-layer decoder with hidden size $$\mathcal{F}(b_i; \mu_k, \gamma_k) = \prod_{j \in D} \mathcal{F}(b_{i,j}; \mu_{k,j}, \gamma_{k,j}), \quad D=\{l,t,r,b\}.$$4, and replaces alternate feed-forward transformer blocks in the decoder with MoE blocks. The backbone still extracts generic multiscale features, and the non-MoE layers still pool shared information across all data. Each expert is placed on a separate GPU in the implementation, so the model keeps roughly the same parameter count as the non-MoE version aside from a small router layer. The classification layer is expanded to accommodate the combined class set, producing only a marginal parameter increase overall.

The routing follows a standard sparse MoE formulation. For input token $$\mathcal{F}(b_i; \mu_k, \gamma_k) = \prod_{j \in D} \mathcal{F}(b_{i,j}; \mu_{k,j}, \gamma_{k,j}), \quad D=\{l,t,r,b\}.$$5, a router with weights $$\mathcal{F}(b_i; \mu_k, \gamma_k) = \prod_{j \in D} \mathcal{F}(b_{i,j}; \mu_{k,j}, \gamma_{k,j}), \quad D=\{l,t,r,b\}.$$6 computes

$$\mathcal{F}(b_i; \mu_k, \gamma_k) = \prod_{j \in D} \mathcal{F}(b_{i,j}; \mu_{k,j}, \gamma_{k,j}), \quad D=\{l,t,r,b\}.$$7

With top-$$\mathcal{F}(b_i; \mu_k, \gamma_k) = \prod_{j \in D} \mathcal{F}(b_{i,j}; \mu_{k,j}, \gamma_{k,j}), \quad D=\{l,t,r,b\}.$$8 routing, the output is

$$\mathcal{F}(b_i; \mu_k, \gamma_k) = \prod_{j \in D} \mathcal{F}(b_{i,j}; \mu_{k,j}, \gamma_{k,j}), \quad D=\{l,t,r,b\}.$$9

DAMEX uses $K$0, following Switch/Tutel-style sparse routing, for throughput benefit with only a small accuracy tradeoff. To avoid overloading a few experts, it uses a load-balancing auxiliary loss from importance and load statistics, but applies this balancing loss only on foreground tokens. The authors state that this improves gradient quality and reduces representation collapse for object detection.

Its distinctive component is the dataset-aware routing objective. The model is given a mapping

$K$1

where each dataset $K$2 is assigned to a target expert $K$3. The router is trained with

$K$4

so that tokens from dataset $K$5 are encouraged to flow to the mapped expert. At inference, the model does not need the dataset label; the trained router is expected to infer the appropriate pathway from the image or token content itself. The paper explicitly contrasts this with older multi-dataset detectors that require explicit dataset or domain knowledge at test time.

The evaluation is on the Universal Object-Detection Benchmark, which mixes 11 heterogeneous datasets including natural images, traffic scenes, aerial surveillance, medical images, stylized cross-domain data, indoor kitchen scenes, and face detection. In the full multi-dataset setting, mixed DINO reaches $K$6 AP, mixed DINO-MoE reaches $K$7 AP, and DAMEX reaches $K$8 AP. The paper reports that DAMEX improves over the non-MoE mixed baseline by $K$9 AP on average, over vanilla MoE by $M^s = \{\pi^s, \mu^s, \gamma^s, p^s, o^s\}$0 AP, and over the previous state of the art from Wang et al. by $M^s = \{\pi^s, \mu^s, \gamma^s, p^s, o^s\}$1 AP. In a few-shot experiment mixing Kitchen with reduced KITTI subsets, DAMEX improves relative to the non-MoE baseline by $M^s = \{\pi^s, \mu^s, \gamma^s, p^s, o^s\}$2 in the 50-shot setting and $M^s = \{\pi^s, \mu^s, \gamma^s, p^s, o^s\}$3 in the 100-shot setting. For Watercolor and Comic, it reaches $M^s = \{\pi^s, \mu^s, \gamma^s, p^s, o^s\}$4 AP on Watercolor and $M^s = \{\pi^s, \mu^s, \gamma^s, p^s, o^s\}$5 AP on Comic, outperforming both mixed DINO and mixed DINO-MoE. The paper also reports that the best performance on a four-dataset mixture is achieved when the number of experts matches the number of datasets; using more experts than datasets does not outperform the matched setup.

The paper provides both qualitative and quantitative evidence against representation collapse. Vanilla DINO-MoE shows poor specialization and mode collapse early in training, whereas DAMEX learns clearer alignment between datasets and experts. A human-prior experiment further shows that random dataset-to-expert mapping still beats vanilla MoE, while a mapping based on domain similarity improves further. A plausible implication is that the gain arises not only from conditional capacity, but from structured specialization aligned with dataset heterogeneity.

4. Full-detector experts and adaptive gating

A more literal model-level mixture uses complete detectors as experts. "YOLO Meets Mixture-of-Experts: Adaptive Expert Routing for Robust Object Detection" (Meiraz et al., 17 Nov 2025) builds an MoE from multiple YOLOv9-T detectors rather than from shared-backbone heads or from test-time ensembling after independent inference. Each expert is a YOLOv9-T branch that processes the image and produces class and bounding-box predictions, and the router allocates responsibility across entire experts at each detection scale.

The architecture follows the multi-scale structure of YOLOv9. It uses three feature levels

$M^s = \{\pi^s, \mu^s, \gamma^s, p^s, o^s\}$6

Each scale produces outputs for bounding box regression and object classification. The box branch predicts distances to box sides discretized into $M^s = \{\pi^s, \mu^s, \gamma^s, p^s, o^s\}$7 bins, typically $M^s = \{\pi^s, \mu^s, \gamma^s, p^s, o^s\}$8, and coordinates are recovered via expectation over predicted probabilities. For expert $M^s = \{\pi^s, \mu^s, \gamma^s, p^s, o^s\}$9 at level $s$0, the intermediate representation is

$s$1

The router is scale-specific and uses expert-derived features. To capture interactions among experts, the model computes

$s$2

where $s$3 is learnable. The router input is

$s$4

and a lightweight CNN with downsampling followed by a fully connected layer produces routing scores

$s$5

Fusion is then

$s$6

for both classification logits and bounding-box logits. To prevent collapse, the total loss is

$s$7

with a load-balancing term inspired by Switch Transformers. The reported setting uses two YOLOv9-T experts and $s$8.

The experiments are on COCO and VisDrone with shared categories Person, Vehicle, Bicycle, and Motorcycle, for 50 epochs using default YOLOv9 hyperparameters and best-model selection by mAP. On COCO test, YOLOv9-T trained on COCO reaches $s$9 mAP and $g=\{g_1,\dots,g_N\}$00 AR, YOLOv9-T trained on COCO+VisDrone reaches $g=\{g_1,\dots,g_N\}$01 mAP and $g=\{g_1,\dots,g_N\}$02 AR, and MoE-T trained on COCO+VisDrone reaches $g=\{g_1,\dots,g_N\}$03 mAP and $g=\{g_1,\dots,g_N\}$04 AR. On VisDrone test, YOLOv9-T trained on VisDrone reaches $g=\{g_1,\dots,g_N\}$05 mAP and $g=\{g_1,\dots,g_N\}$06 AR, YOLOv9-T trained on COCO+VisDrone reaches $g=\{g_1,\dots,g_N\}$07 mAP and $g=\{g_1,\dots,g_N\}$08 AR, and MoE-T trained on COCO+VisDrone reaches $g=\{g_1,\dots,g_N\}$09 mAP and $g=\{g_1,\dots,g_N\}$10 AR. The paper attributes the gains to specialization, scale-aware routing, and balancing regularization.

"Domain-Specialized Object Detection via Model-Level Mixtures of Experts" (Pavlitska et al., 20 Apr 2026) studies a related detector-level MoE under explicit domain partitioning. Its experts are full YOLO-based detectors trained on semantically disjoint subsets of BDD100K: one daytime expert and one nighttime expert. The gating network $g=\{g_1,\dots,g_N\}$11 consumes selected intermediate feature maps $g=\{g_1,\dots,g_N\}$12 and predicts

$g=\{g_1,\dots,g_N\}$13

If $g=\{g_1,\dots,g_N\}$14 denotes the raw output of expert $g=\{g_1,\dots,g_N\}$15, then

$g=\{g_1,\dots,g_N\}$16

The weighted predictions are decoded with a shared anchor configuration and then merged by NMS or NMW. The paper explicitly motivates pre-decoding weighting as a way to preserve expert identity longer than naive output concatenation.

Several gate designs are evaluated: AvgPool → FC(2) → Softmax, AvgPool → FC(512) → FC(2) → Softmax, Conv(1×1) → AvgPool → FC(512) → FC(2) → Softmax, and Conv(1×1) → BN → Conv(3×3) → BN → AvgPool → FC(512) → FC(2) → Softmax. The best-performing gate is the one-conv, two-FC version. The paper also studies single image-level gates, spatial gates, and class-wise gates; the single image-level gate performs best.

The detector backbones include YOLOv7 and YOLOv9, with large and medium variants, both from scratch and pretrained on COCO. Training uses 200 epochs, confidence threshold $g=\{g_1,\dots,g_N\}$17, IoU threshold $g=\{g_1,\dots,g_N\}$18, and mAP50 evaluation. Anchors for YOLOv7 are optimized on combined training data, and YOLOv9 inference uses the primary detection head. The paper studies balancing losses based on importance, KL divergence, entropy, and sample-wise entropy regularization, and emphasizes that sample-wise entropy is the best-performing balancing objective. It also examines domain-aware gate training with cross-entropy on domain labels and finds it weaker overall than unsupervised routing.

The reported results are deliberately nuanced. Each expert performs best on its own domain, the baseline trained on full daytime+nighttime data is usually strongest overall, the MoE consistently beats the fixed ensemble, and the MoE is usually close to, but not always better than, the strongest single baseline. For YOLOv7x pretrained on COCO, the baseline all-data model reaches $g=\{g_1,\dots,g_N\}$19 mAP, the ensemble reaches $g=\{g_1,\dots,g_N\}$20 mAP, and the MoE reaches $g=\{g_1,\dots,g_N\}$21 mAP. For YOLOv9-E pretrained on COCO, the baseline reaches $g=\{g_1,\dots,g_N\}$22 mAP, the ensemble $g=\{g_1,\dots,g_N\}$23 mAP, and the MoE $g=\{g_1,\dots,g_N\}$24 mAP. Among ensemble fusion baselines on YOLOv7x, NMS gives $g=\{g_1,\dots,g_N\}$25, WBF $g=\{g_1,\dots,g_N\}$26, NMW $g=\{g_1,\dots,g_N\}$27, and Soft-NMS $g=\{g_1,\dots,g_N\}$28; with mAP re-weighting, NMS gives $g=\{g_1,\dots,g_N\}$29, WBF $g=\{g_1,\dots,g_N\}$30, NMW $g=\{g_1,\dots,g_N\}$31, and Soft-NMS $g=\{g_1,\dots,g_N\}$32, making NMW with mAP re-weighting the best ensemble baseline. Routing analysis shows higher gate weight on the daytime expert for daytime images, higher weight on the nighttime expert for nighttime images, skew toward the daytime expert on dawn/dusk, and more balanced routing on undefined images, which the paper interprets as uncertainty or OOD behavior.

These detector-level MoEs show that expert specialization can be organized either implicitly, as in scale-wise routing over YOLOv9-T experts, or explicitly, as in day/night specialization on BDD100K. A plausible implication is that full-detector MoEs become attractive when heterogeneity is strong enough that sharing only one dense pathway is limiting, but not so strong that expert outputs become impossible to reconcile.

5. Calibration, reranking, and the boundary with late fusion

The literature also contains methods that combine detectors without embedding the mixture inside a jointly routed detection model. These methods are important because they clarify what model-level mixtures are not, and because several later works identify calibration and duplicate suppression as central difficulties.

"Detect2Rank: Combining Object Detectors Using Learning to Rank" (Karaoglu et al., 2014) combines three publicly available detectors—DPM, CN, and EES—by constructing a 34-dimensional contextual feature vector for each detection and learning a ranking function

$g=\{g_1,\dots,g_N\}$33

Its feature groups are detector-detector context $g=\{g_1,\dots,g_N\}$34, object likelihood $g=\{g_1,\dots,g_N\}$35, and object-object context $g=\{g_1,\dots,g_N\}$36. The workflow is to run DPM, CN, and EES independently, collect all detections, compute contextual features, train a learning-to-rank model, rescore detections, merge them into one list, and apply NMS after ranking with a $g=\{g_1,\dots,g_N\}$37 overlap threshold. The best-performing pointwise method is logistic regression. On VOC07, baseline mAP values are $g=\{g_1,\dots,g_N\}$38 for DPM, $g=\{g_1,\dots,g_N\}$39 for CN, and $g=\{g_1,\dots,g_N\}$40 for EES, while the combined detector M-All reaches $g=\{g_1,\dots,g_N\}$41 mAP; on VOC10, the paper reports improvements of $g=\{g_1,\dots,g_N\}$42 over DPM, $g=\{g_1,\dots,g_N\}$43 over CN, and $g=\{g_1,\dots,g_N\}$44 over EES. Detect2Rank is therefore a strong detector-combination framework, but its mixture occurs through learned reranking and post-hoc suppression rather than through internal expert routing.

"Dynamic Belief Fusion for Combining Multiple Object Detectors" (Lee et al., 2022) makes this distinction explicit. It is described as a score-level late fusion method rather than a model-level mixture. DBF builds a confidence model from each detector’s precision-recall relationship, assigns belief mass over hypotheses $g=\{g_1,\dots,g_N\}$45, $g=\{g_1,\dots,g_N\}$46, and the intermediate state $g=\{g_1,\dots,g_N\}$47, and fuses belief distributions with Dempster’s combination rule. Its dynamic assignment is

$g=\{g_1,\dots,g_N\}$48

and the theoretical detector curve is

$g=\{g_1,\dots,g_N\}$49

The paper emphasizes that the intermediate hypothesis $g=\{g_1,\dots,g_N\}$50 is essential for ambiguity modeling. DBF outperforms individual detectors and fusion baselines on ARL, VOC07, and VOC12: on ARL, the best individual detector DPM reaches mAP $g=\{g_1,\dots,g_N\}$51 and DBF reaches $g=\{g_1,\dots,g_N\}$52; on VOC07, DBF reaches $g=\{g_1,\dots,g_N\}$53 in the heterogeneous-detector setting and $g=\{g_1,\dots,g_N\}$54 with CNN-only detectors; on VOC12, DBF reaches $g=\{g_1,\dots,g_N\}$55. The paper’s own conclusion is that DBF should be understood as a late-fusion precursor or alternative to model-level mixtures, not as one itself.

"MoCaE: Mixture of Calibrated Experts Significantly Improves Object Detection" (Oksuz et al., 2023) occupies an intermediate position. It does not train a new jointly routed detector from scratch; instead, it reuses already-trained detectors, calibrates their scores so that confidence reflects localization quality, and then fuses predictions with refined postprocessing. Its calibration criterion ties confidence to IoU: $g=\{g_1,\dots,g_N\}$56 The practical calibration error metric is LaECE, with class-agnostic isotonic regression reported as the best tradeoff. Only 500 calibration images are usually enough. After calibration, MoCaE concatenates detections and applies Refining NMS, which combines Soft NMS and Score Voting. The paper argues that naïve ensembles or “Vanilla MoE” degrade because the most confident detector dominates the merged output regardless of whether it is actually the best detector on that image. In ablations, uncalibrated MoE performs about like a single model and sometimes worse, whereas late calibration plus calibrated Soft NMS plus Score Voting yields the best COCO mini-test result of $g=\{g_1,\dots,g_N\}$57 AP, improving over the best single model’s $g=\{g_1,\dots,g_N\}$58 AP. The broader results include up to $g=\{g_1,\dots,g_N\}$59 AP improvement on COCO and LVIS tasks, $g=\{g_1,\dots,g_N\}$60 AP and $g=\{g_1,\dots,g_N\}$61 AP on COCO test-dev in different settings, $g=\{g_1,\dots,g_N\}$62 mask AP on LVIS, $g=\{g_1,\dots,g_N\}$63 AP50 on DOTA, and improvements on open-vocabulary and robustness benchmarks. A common misconception is that any detector ensemble is already a useful MoE; MoCaE directly argues against this by showing that detector miscalibration can make a naïve mixture worse than its components.

6. Recurring findings, interpretability, and limitations

Several empirical regularities recur across these papers. First, specialization is the dominant mechanism. In Mixture Dense Regression, components specialize by object scale rather than category; in DAMEX, experts align with datasets or groups of similar datasets; in the YOLOv9-T MoE, specialization is described as emerging in feature space rather than through manually assigned roles; in the BDD100K MoE, the gate naturally learns daytime versus nighttime routing even without domain-aware supervision (Varamesh et al., 2019, Jain et al., 2023, Meiraz et al., 17 Nov 2025, Pavlitska et al., 20 Apr 2026). This suggests that model-level mixtures become most informative when there is a latent partition—scale, dataset, or domain—that a single homogeneous pathway does not represent efficiently.

Second, balancing or anti-collapse regularization is consistently important. D-RMM uses maximum component maximization to prevent several components from explaining the same object. DAMEX uses load balancing only on foreground tokens and reports reduced representation collapse. The YOLOv9-T detector-level MoE adds a Switch-style load-balancing loss. The BDD100K MoE studies importance, KL-divergence, entropy, and sample-wise entropy balancing, with sample-wise entropy performing best. Across formulations, the regularizer is not an auxiliary convenience; it is part of the definition of a workable mixture because expert or component redundancy otherwise persists (Yoo et al., 2022, Jain et al., 2023, Meiraz et al., 17 Nov 2025, Pavlitska et al., 20 Apr 2026).

Third, duplicate handling and confidence reliability are central detection-specific issues. D-RMM builds deduplication into the training objective via MCM and reports improved confidence reliability and Expected Calibration Error. MoCaE argues that cross-detector miscalibration is the primary reason naïve mixtures fail, because NMS is then driven by incomparable scores. The BDD100K MoE notes that object detection is difficult for MoEs because the outputs are dense, structured, and variable-length, so weighting must happen before decoding and suppression if expert attribution is to remain meaningful (Yoo et al., 2022, Oksuz et al., 2023, Pavlitska et al., 20 Apr 2026). A plausible implication is that object detection mixtures differ from classification MoEs less in routing mechanics than in how they must manage structured outputs and postprocessing.

Fourth, gains are strongest in heterogeneous settings. DAMEX reports consistent gains under limited data, disparate domains, and divergent label sets. The YOLOv9-T MoE improves jointly trained performance on both COCO and VisDrone. The BDD100K MoE shows complementary expert behavior across day, night, dawn/dusk, and undefined subsets. MoCaE gains are largest when experts are diverse but not wildly mismatched in quality. These results collectively support the view that model-level mixtures are especially relevant when the training or deployment distribution is a mixture of domains rather than a single homogeneous corpus (Jain et al., 2023, Meiraz et al., 17 Nov 2025, Pavlitska et al., 20 Apr 2026, Oksuz et al., 2023).

The limitations are equally consistent. Model-level MoEs do not always beat the strongest fully joint baseline: the BDD100K study states this explicitly, noting that the baseline trained on all data can still be better because it sees the entire distribution (Pavlitska et al., 20 Apr 2026). More components are not always better: in Mixture Dense Regression, $g=\{g_1,\dots,g_N\}$64 yields cleaner scale separation than $g=\{g_1,\dots,g_N\}$65 but not higher AP (Varamesh et al., 2019). More experts than datasets do not improve DAMEX beyond the matched setting (Jain et al., 2023). Domain-aware gate training can be weaker than unsupervised routing in the BDD100K MoE (Pavlitska et al., 20 Apr 2026). MoCaE also notes that when one detector is much stronger than the others, the mixture may not beat the best single detector because low-capacity calibrators cannot fully equalize very different models without damaging ranking (Oksuz et al., 2023).

The resulting picture is neither that mixtures are uniformly superior nor that they are merely ensembles under a new name. The evidence instead supports a narrower claim: model-level mixtures of object detectors are effective when the mixture mechanism is aligned with a concrete source of heterogeneity, when duplicate suppression or score calibration is treated as a first-class modeling problem, and when expert usage is regularized strongly enough to prevent collapse. Under those conditions, the literature shows three viable instantiations—probabilistic internal mixtures, sparse expert routing in shared detectors, and gates over full detector experts—each with distinct trade-offs in parameter sharing, interpretability, and dependence on postprocessing.