acceptodds
Withdrawn from ICLR 2027

MoRE-DETR: Adaptive Recursive Reasoning with Vision-Language Model for Enhancing Remote Sensing Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Remote sensing object detection suffers from severe foreground-background imbalance and substantial variations in object scale and density, making foreground tokens differ greatly in detection difficulty. Existing DETR-like detectors refine queries uniformly and thus under-refine hard objects, while VLM-assisted variants either retain the VLM during inference or generate coordinates as text with weak geometric supervision. To address these issues, we propose MoRE-DETR, a DETR-based framework that couples adaptive token computation with VLM-assisted training on a shared token representation. Specifically, an Adaptive Recursive Token Routing (ARTR) module employs an uncertainty-aware Mixture-of-Recursions router to allocate recursive depth according to token difficulty, producing difficulty-adaptive foreground tokens. These refined tokens are fed to a Visual-Language Aligned Supervision (VLAS) module, which projects them into the semantic space of a multimodal large language model and directly regresses continuous bounding boxes from its hidden states, thereby injecting explicit geometric supervision back into the token representation. VLAS is used only during training and introduces no additional inference overhead. Extensive experiments on multiple remote sensing benchmarks demonstrate state-of-the-art performance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
object detection, remote sensing, computer vision

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_twBJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes MoRE-DETR, a remote sensing object detector that combines adaptive recursive token refinement with vision-language model-assisted supervision. The ARTR module allocates recursive computation according to token importance and classification uncertainty, while the VLAS module introduces auxiliary semantic and continuous bounding-box supervision through a pretrained VLM during training.

Critical strengths
  1. The paper addresses the uneven computational demands of remote sensing objects through difficulty-aware recursive token refinement.

  2. The training-only VLM supervision strategy provides additional semantic and geometric guidance without requiring the VLM during inference.

Critical weaknesses
  1. The methodological novelty is somewhat limited, as ARTR largely builds upon the existing foreground token selection and Mixture-of-Recursions mechanisms, while VLAS follows a conventional auxiliary supervision paradigm. The paper does not sufficiently establish a distinctive technical contribution beyond their integration.

  2. The uncertainty-aware routing strategy lacks direct evidence that high-entropy tokens correspond to genuinely difficult objects. Classification entropy may also be high for background regions or poorly calibrated predictions, and the paper does not analyze routing decisions across object scales, detection errors, or recursion depths.

  3. The effectiveness of VLM-specific knowledge transfer remains unclear. Although the auxiliary box regression loss improves performance, the experiments do not compare against an equally parameterized visual-only auxiliary head or a randomly initialized language model, making it difficult to isolate the contribution of pretrained vision-language knowledge.

  4. The claimed coupling between ARTR and VLAS is insufficiently validated. The current ablations demonstrate improvements from combining the two modules but do not establish whether VLAS preferentially benefits recursively refined difficult tokens or whether the gains arise from independent feature enhancement and auxiliary supervision.

Questions for authors

see the Weaknesses

Reviewer_qH2tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes MoRE-DETR, a new DETR-based framework that couples adaptive token computation with VLM-assisted training for remote sensing object detection. The paper mainly applies the Mixture-of-Recursions (MoR) paradigm into RSOD and makes a series of adaptive changes.. The Adaptive Recursive Token Routing (ARTR) module allocate recursive depth according to foreground token difficulty. The Vision‑Language Aligned Supervision (VLAS) module provides auxiliary training supervision from a pre‑trained multimodal VLM. Experiments are conducted on three standard RSOD datasets, DOTA‑1.0, AI‑TOD and VisDrone. The improvement is marginal comparing with previous state-of-the-art methods.

Critical strengths

Applying MoR intro remote sensing object detection is somewhat interesting.

Critical weaknesses
  1. The experiments can hardly support the effectiveness of the proposed method. The improvement is marginal compared to previous state-of-the-art methods. The proposed MoRE-DETR only achieves 0.4, 0.2, and 0.5 AP gains on DOTA-1.0, VisDrone, and AI-TOD datasets, respectively. Such small margins may fall within the range of common training variance and hyper‑parameter tuning space for remote sensing detectors. The paper also claims that "MoRE-DETR improves small and difficult object detection in remote sensing imagery". However, compared with DetectoRS/VLM4RSDet on VisDrone dataset, MoRE-DETR achieves only 0.1 AP$_{vt}$ gain and even drops 0.2 AP$_t$.

  2. In section 4.2, Quantitative comparison (3) "Competitive performance on tiny objects. MoRE-DETR achieves the same AP$_{vt}$ as VLM4RSDet on AI-TOD, with AP$_t$ only 0.2% lower." Where are the results? There is no VLM4RSDet result on DOTA and AI-TOD.

Questions for authors

Since the paper claims to address small and difficult object detection in remote sensing imagery, it should present direct comparison on small objects.

Reviewer_BJZsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses closed-set remote-sensing object detection (RSOD) with DETR-style detectors. Its premise is that foreground tokens differ substantially in detection difficulty, that existing detectors refine all queries uniformly, and that VLM-assisted detectors either keep the VLM at inference (extra cost) or regress boxes as text (weak geometric supervision). Two modules are proposed on a shared token representation. ARTR is an uncertainty-aware mixture-of-recursions router (expert-choice routing, R = 2 recursions, N_r = 2 shared blocks) that allocates recursive depth per token according to an MLP-based uncertainty score, producing difficulty-adaptive foreground tokens. VLAS projects those tokens into the semantic space of InternVL3.5-1B (Qwen3-0.6B language model) and regresses continuous bounding boxes from the VLM's hidden states; it is used only during training. The detector is built on Focus-DETR and evaluated on DOTA-1.0, VisDrone and AI-TOD, with ablations over routing strategies and VLM-supervision strategies. Reported results are 73.0 mAP50 on DOTA-1.0, 31.6 mAP on VisDrone and 56.6 mAP50 on AI-TOD, with 46.24M parameters and 117.8G FLOPs

Critical strengths

S1. A lean, deployable design with a genuinely large improvement over its own baseline.MoRE-DETR adds nothing at inference (the VLM is discarded), uses 46.24M parameters and 117.8G FLOPs, and improves VisDrone mAP from 23.5 (Focus-DETR) to 31.6. If the numbers hold, the practical message — VLM supervision can be cashed out entirely at training time — is valuable for the RS detection community, which is sensitive to deployment cost. S2. The VLM-supervision ablation is more granular than is usual in this genre. Four supervision strategies with matched teacher and budget (frozen, CE only, CE+box, unfreezing the last two FFN blocks) isolate a clean effect: continuous box regression from adapted hidden states beats text-level CE by 0.9 mAP. This is a concrete, transferable recipe rather than an architectural claim.

Critical weaknesses

W1. The margins over the state of the art are within noise, and the paper neither measures nor reports that noise. DOTA-1.0: 73.0 vs 72.6 (+0.4 mAP50); VisDrone: 31.6 vs 31.4 (+0.2 mAP, i.e. 0.6% relative); AI-TOD: 56.6 vs 56.1 (+0.5 mAP50). There are no seeds, no confidence intervals, and no repeated runs anywhere in the paper, yet "state-of-the-art" and "consistently improves" are the abstract's and conclusion's main claims. Worse, the DOTA average conceals large per-class regressions: against SKYDET-CT, MoRE-DETR loses on seven of fifteen classes, including BR 57.7 vs 81.5 (−23.8), BC 61.0 vs 83.8 (−22.8) and HA 82.4 vs 94.2 (−11.8), while the average is recovered almost entirely by TC (+32.8) and ST (+24.5). A method that trades −24 AP on bridges and −23 AP on basketball courts for +33 AP on tennis courts is not uniformly better, and the paper does not discuss this at all. On AI-TOD it is also below SKYDET-T on mAP75 (22.4 vs 23.5). W2. The ablation cannot be reconciled with the main results, so the per-module attributions are not established. Table 4's full configuration (ARTR with MLP+Unc., plus VLAS) is 29.2 mAP / 50.5 mAP50, and Table 5's best configuration is 29.6 / 51.2 — yet Table 2 reports 31.6 mAP / 53.7 mAP50 for the same dataset, metric and model. A 2.0–2.4 point gap between "the ablation's full model" and "the paper's model" is not a rounding issue; either the ablation used a different schedule, resolution or evaluation protocol (which must be stated, because then the +0.7/+1.8/+5.7 attributions are not comparable to the headline), or one of the two tables is wrong. Relatedly, the combined gain (+5.7) is more than twice the sum of the individual gains (+0.7 ARTR, +1.8 VLAS), implying a 3.2-point interaction that is asserted but never analysed; with one row of a table as evidence, "the two modules are complementary" is not supported. W3. The VLM-supervision comparison is self-referential. Table 5 varies only the loss (CE vs box) and how much of the same InternVL is unfrozen. What is missing is exactly what would show that the geometric supervision matters rather than "more supervision from a larger model": (a) distillation from frozen VLM features (MSE/cosine) with the same teacher, (b) the text-embedding alignment used by the VLM-assisted RS detectors the paper positions against (e.g. VLM4RSDet), and (c) the same VLM box loss applied to a non-adaptive, uniformly refined token set, which would separate the VLAS effect from the ARTR interaction. Without (a)–(c), the paper's claim that adaptive recursion and VLM supervision are complementary is an architectural assertion rather than a measured result.

Questions for authors
  1. Please reconcile Tables 4 and 5 (29.2 / 29.6 mAP) with Table 2 (31.6 mAP) on VisDrone: same schedule, input resolution, and evaluation protocol? If not, please report the ablation under the main setting.
  2. Please report mean ± std over at least three seeds for DOTA-1.0, VisDrone and AI-TOD, and add a per-class delta table for DOTA so that the BR and HA regressions are visible.
  3. Can you provide controls (a)–(c) in W3, in particular the VLM box loss on a uniformly refined token set? If the gain disappears, does the paper still claim ARTR is necessary?
  4. What is the measured latency and throughput of MoRE-DETR vs Focus-DETR and RT-DETRv4 on the same GPU? Does the router's early exit reduce wall-clock time, or only FLOPs?
  5. Why is the joint gain (+5.7) so much larger than the sum of the individual gains (+2.5)? Which mechanism produces the interaction?
  6. How is token difficulty defined at inference, and how does the uncertainty threshold behave on scenes with different object densities (e.g. DOTA vs VisDrone)?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.