acceptodds
Withdrawn from ICLR 2027

FlexDepth++: Towards a Unified Self-Supervised Multi-Frame Monocular Depth Estimation Framework for Driving and Low-Altitude Aerial Scenes

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Self-supervised multi-frame monocular depth estimation has achieved remarkable progress in autonomous driving by leveraging temporal geometric consistency from unlabeled videos. However, existing methods are mostly developed and evaluated in structured driving environments, making it non-trivial to extend them to low-altitude aerial scenes. Compared with vehicle-mounted driving videos, low-altitude aerial videos exhibit broader and more uneven depth distributions, more complex camera motion, and distinct scene structures, which undermine the reliability of temporal geometric matching. Since cost volumes are the central representation for aggregating cross-frame matching cues, building reliable and adaptable cost-volume representations is crucial for a unified multi-frame depth estimation framework across these heterogeneous scenarios. In this paper, we propose FlexDepth++, a unified self-supervised multi-frame depth estimation framework for driving and low-altitude aerial scenes. FlexDepth++ is centered on hierarchical cost-volume refinement, which follows a construction-then-refinement paradigm with two key components: (1) multi-resolution cost-volume construction, which builds cost volumes from complementary feature levels to capture coarse geometric matching over broad depth ranges while preserving fine local correspondence around high-frequency structures; and (2) consistency-guided refinement, which leverages a training-derived consistency mask as an auxiliary spatial prior to augment cost-volume representations and implicitly guide the model toward geometrically reliable regions. Together, these components enhance cost-volume representations from both scale and reliability perspectives, enabling richer and more reliable temporal geometric cues for robust depth prediction across both driving and low-altitude aerial scenarios. Experiments on UAVScene, UAVLight, KITTI, and Cityscapes show that FlexDepth++ consistently improves self-supervised depth estimation across both low-altitude aerial and autonomous driving domains. To the best of our knowledge, FlexDepth++ is among the first self-supervised multi-frame depth estimation frameworks systematically designed and validated for both driving and low-altitude aerial scenes, demonstrating the effectiveness of hierarchical cost-volume refinement under different camera motion patterns, depth distributions, and scene structures.

Primary area:
applications to robotics, autonomy, planning
Keywords:
3D Perception, Unmanned Aerial Vehicle (UAV)

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ZmEcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Proposes FlexDepth++, a self-supervised multi-frame monocular depth estimation framework for driving and UAV scenes. Combines multi-resolution cost-volume construction with stability-guided refinement to improve temporal matching. Achieves strong depth estimation results on four benchmarks, including KITTI and UAVScenes.

Critical strengths
  1. Strong experimental results
  • Achieves competitive or state-of-the-art performance across four driving and UAV benchmarks.
  1. Practical problem
  • Addresses temporal matching challenges across different camera motions and depth distributions.
  • Ablations support the effectiveness of the proposed modules.
Critical weaknesses
  1. Multi-resolution cost-volume fusion and stability-based refinement are relatively straightforward extensions of existing approaches. The key technical contribution is not sufficiently clear.
  2. Models are trained separately for each dataset with different configurations. No cross-domain transfer or joint-training experiments to support the claimed generalization.
  3. Additional cost volumes and refinement introduce computational overhead, but runtime, FLOPs, and memory costs are not reported.
Questions for authors
  1. What is the main methodological novelty beyond existing multi-scale cost-volume methods?
  2. Can a single trained model generalize across driving and UAV domains without retraining?
  3. What are the inference latency, FLOPs, and memory overhead compared with FlexDepth?
  4. How much does SCVR improve performance beyond MCVC, particularly on KITTI?
Reviewer_Jf3pOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors propose a self-supervised framework named FlexDepth++ for monocular depth estimation in both driving and low-altitude aerial scenes. To improve depth estimation performance, the authors introduce two modules: Multi-resolution Cost-Volume Construction and Stability-guided Cost-Volume Refinement. Experimental results show that FlexDepth++ achieves state-of-the-art performance in UAVScenes, UAVLight, KITTI, and Cityscapes.

Critical strengths

The authors construct a multi-resolution cost volume to aggregate temporal information and improve the method’s performance, which I consider useful. The FlexDepth++ framework follows a common design for depth estimation in driving scenes. The experiments are comprehensive, covering both UAV and driving scenes, and the results show that FlexDepth++ achieves SOTA performance.

Critical weaknesses

I do not quite understand why the authors want to using FlexDepth++ to both UAV and driving scenes, and I would appreciate a clearer explanation of the motivation and value of doing so. Furthermore, the experimental settings show that the model is trained and tested separately in these two scenes, without combining the two type of data. Therefore, it is unclear to me what additional benefits come from studying both settings in the same paper.

Questions for authors

See the Weaknesses above.

Reviewer_S3C2OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

FlexDepth++ extends the self-supervised ManyDepth and FlexDepth pipeline for multi-frame monocular depth estimation to both driving and low-altitude UAV scenes. Given target and source frames, a pose network estimates relative motion, source features are warped over 96 candidate depths, and L1 cost volumes are built at two feature resolutions. Multi-resolution Cost-Volume Construction concatenates and fuses these volumes with a small 3×3 convolution stack; Stability-guided Cost-Volume Refinement compares disparity maps from early and late snapshots of a frozen first-stage model, thresholds the discrepancy into a binary mask, and uses it as an auxiliary spatial prior on the fused cost volume. The enhanced cost volume is injected into a FlexDepth decoder over a YOLOv11 encoder. Training follows Monodepth2 view synthesis, auto-masking, edge smoothness, and the ManyDepth monocular-versus-cost-volume consistency loss. Results are reported on UAVScenes, UAVLight, KITTI Eigen, and Cityscapes, with Abs Rel of 0.071 and 0.046 on the two UAV sets, 0.082 on KITTI, and 0.083 on Cityscapes, plus ablations on KITTI and UAVScenes. The paper addresses an underexplored setting and reports broad benchmark numbers, but the two proposed modules are light combinations of existing multi-resolution cost-volume and training-dynamics mask ideas, the stability prior is under-specified, baselines mix input modalities, and no efficiency measure, joint-domain training, or cross-domain transfer experiment is provided.

Critical strengths

1.The pipeline is straightforward and reproducible. It uses the standard ManyDepth warping and L1 matching, standard Monodepth2 losses, fixed 96 depth bins, and reports per-dataset learning rates, batch size, epoch count, depth ranges, and consistency weights in the implementation section.

2.Both modules are ablated on a driving and a UAV set. On UAVScenes the gains over the listed baseline are large, and the full model improves Sq Rel and threshold accuracy alongside Abs Rel, rather than improving only one metric.

3.Qualitative figures compare against zero-shot Depth Anything V2 and ManyDepth2 and discuss depth-range distortion in aerial scenes, a failure mode the self-supervised temporal approach avoids more consistently than zero-shot foundation priors.

Critical weaknesses

1.Conceptual novelty is incremental. The encoder is YOLOv11, the decoder and initialization are FlexDepth, pose and photometric losses follow Monodepth2 and ManyDepth, and the L1 warped cost volume follows ManyDepth. MCVC reuses the established multi-resolution cost-volume and feature-pyramid aggregation idea already present in stereo and multi-frame depth literature; the only change is two levels concatenated and passed through a fusion convolution. SCVR adapts an existing training-dynamics discrepancy mask, cited as inspired by Zhang et al. 2025, to a cost volume. The paper does not introduce a new matching principle, uncertainty model, or geometric constraint that separates this module from prior uncertainty maps and self-inhibition masks.

2.The ablation evidence is weak on driving data. On KITTI, the baseline reaches 0.084 Abs Rel, MCVC 0.082, SCVR 0.082, and the full model 0.082, with δ1 moving from 0.923 to 0.926; these margins are small without seed variance or confidence intervals.

3.Baselines are not matched by input modality. Several compared methods rely on optical flow or semantic labels, while FlexDepth++ uses monocular video only; these entries should be separated or given equivalent inputs. MoGe and MoGe-2, identified in related work as representative open-domain aerial geometry models, are absent from the quantitative tables, and several recent 2025–2026 self-supervised multi-frame depth methods are not included on KITTI and Cityscapes. Evaluation details such as the Eigen crop, median versus mean scaling, and depth versus disparity normalization are not restated in the experimental section.

Questions for authors

1.The pipeline reuses a YOLOv11 encoder, a FlexDepth decoder and checkpoint, Monodepth2 losses, and the ManyDepth L1 cost volume, while MCVC and SCVR appear as light additions on top. Could you state the single core novel contribution of this paper and describe its formal difference from existing multi-resolution cost-volume designs and existing training-dynamics uncertainty masks? The review needs to judge whether the paper introduces a new mechanism rather than an engineering combination of established components.

2.The paper claims a unified framework across driving and low-altitude UAV scenes. Beyond sharing an architecture, what general methodological innovation transfers across the two domains? Please clarify whether this is a jointly trained cross-domain model, and identify what evidence would establish a framework-level novelty rather than applying an existing pipeline as-is to two datasets.

3.Could you clarify what evidence shows the reported gains come from the proposed design itself rather than the FlexDepth initialization, the temporal matching recipe, or hyperparameter tuning? Please situate the proposed modules against recent self-supervised multi-frame depth and uncertainty-mask methods so the incremental contribution and novelty relative to the state of the art are clear.

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

Related papers

Discussion (0)

Sign in to comment.