PaperlayerView exact revision on arXiv

Source-linked AI summary

GTR: Gated Token Recurrence for Efficient Dense Prediction

Zhe Feng, Longfei Liu, Wei Liu, Kai Chen, Jiangang Kong, Wei Zhou, Yifeng Qian, Dexiong Chen, Xuanlong Yu, Xi Shen

arXiv:2609.26590v2cs.CVcs.LG

TL;DR

Quadratic softmax-attention costs make high-resolution dense prediction increasingly expensive, particularly on edge devices. GTR combines gated recurrence, directional scanning, and local spatial mixing, and distills from a detection-specialized teacher using final patch-token alignment. Across six dense-prediction tasks, it reports a strong accuracy–efficiency trade-off, including 58.9 box AP for Objects365-pretrained GTR-L on COCO val2017.

  • Problem

    Quadratic softmax-attention complexity makes high-resolution vision processing increasingly expensive, especially on resource-constrained edge devices.

    Source: §1 · p. 2

  • Method

    GTR combines gated linear attention, four-directional scanning, and Spatial SwiGLU, and distills final patch tokens against a frozen detection-specialized DINOv3 teacher.

    Source: §1 · p. 2

  • Results

    Across object detection and five additional dense-prediction tasks, GTR reports a strong accuracy–efficiency trade-off; Objects365-pretrained GTR-L achieves 58.9 box AP on COCO val2017.

    Source: §1 · p. 2, §4 · p. 9

  • Takeaways & Limitations

    The reported results support linear attention as an efficient alternative to softmax attention for dense visual prediction, including automotive edge deployment.

    Source: §4 · p. 9

Abstract

from arXiv · show

Self-attention-based vision backbones perform well on dense prediction, but the quadratic computational cost of global softmax attention limits their efficiency as image resolution increases. We introduce Gated Token Recurrence (GTR), a softmax-free recurrent vision backbone that combines gated linear attention, alternating spatial scan directions, and spatially enhanced SwiGLU blocks. GTR is distilled from a detection-specialized DINOv3 teacher using only final-layer patch-token alignment through a linear projection and squared $\ell_2$ loss, without masked-token prediction or intermediate-layer supervision. With Objects365 detector pre-training, GTR-L achieves 58.9 box AP on COCO \texttt{val2017} with 1.908\,ms median batch-one latency under compiled FP16 execution on an RTX~4090. The same backbone also transfers to instance segmentation, pose estimation, oriented detection, semantic segmentation, and monocular depth estimation. In an isolated kernel benchmark, our specialized chunkwise CUDA operator is $4.0\times$ faster than FLA v0.5.0 at 1.6K tokens on RTX~4090. TensorRT deployment on DRIVE AGX Thor achieves 2.282--8.769\,ms median batch-one latency across the evaluated models. These results show that recurrent token mixing can provide an efficient alternative to global softmax attention for high-resolution dense prediction and edge deployment. Project page: https://intellindust-ai-lab.github.io/projects/GTR/

1 INTRODUCTION

Quadratic softmax attention makes high-resolution dense prediction costly, particularly on edge devices, and a general linear-attention backbone that is accurate across tasks and practically efficient remains challenging. GTR combines gated recurrence, directional scanning, and local spatial mixing, uses final-token distillation, and reports results across six tasks alongside low-latency edge deployment.

1 INTRODUCTION

  • Quadratic softmax-attention cost makes high-resolution processing increasingly expensive, especially on resource-constrained edge devices.The paper frames a general-purpose linear-attention backbone that maintains accuracy across diverse dense-prediction tasks and translates theoretical efficiency into practical edge-hardware gains as the open challenge.Source: §1 · p. 2
  • 58.9 box AP on COCO val2017 at 1.908 ms median latency is reported for Objects365-pretrained GTR-L under compiled FP16 batch-one RTX 4090 inference.The same backbone transfers to six dense-prediction tasks across four model scales.Source: §1 · p. 2
  • A specialized chunkwise CUDA operator is 4.0× faster than FLA v0.5.0 at 1.6K tokens, and TensorRT deployment on DRIVE AGX Thor reports 2.282–8.769 ms median FP16 batch-one latency.The operator comparison uses FP16 and CUDA Graph execution on an RTX 4090.Source: §1 · p. 2
  • GTR combines gated linear recurrence, four-direction scanning, and Spatial SwiGLU to model global context and mix neighboring patches while retaining linear complexity.The backbone uses a single final-patch-token distillation objective with a frozen detection-specialized DINOv3 teacher, without masking or intermediate supervision.Source: §1 · p. 2
  • The paper presents GTR as a single-scale linear-attention backbone for efficient dense prediction, combining four-directional scanning with Spatial SwiGLU for complementary global and local modeling.Its broader evaluation covers object-level and pixel-level tasks using the same backbone design with task-specific adaptation.Source: §1 · p. 2

2 METHOD

GTR maintains a single-scale patch grid through recurrent backbone blocks, combining gated linear attention with directional scans and local spatial mixing. Final-token distillation aligns the student to a frozen detection-specialized teacher, while chunkwise inference parallelizes much of the recurrence; the isolated operator benchmark reports speedups over FLA v0.5.0.

2.1 GTR BACKBONE

  • The backbone maintains a stride-16 patch grid through 12 blocks, using GLA for long-range context, key-only gated recurrence, and Spatial SwiGLU’s depthwise convolution for local mixing.Blocks cycle through four scan directions; detection features from blocks 4, 8, and 12 feed a multi-scale neck and DETR-style decoder.Source: §2 · p. 3, §2.1 · pp. 3–4

2.2 DISTILLATION

  • The distillation pipeline aligns final student patch representations with a frozen detection-specialized DINOv3 teacher using a learned linear projection and squared ℓ2 loss.The loss averages over patch positions, excludes non-patch tokens, and uses no masking or intermediate-layer matching.Source: Figure 3 · p. 4, §2.2 · pp. 4–5

2.3 HARDWARE-AWARE CHUNKWISE INFERENCE

  • 2.6×, 4.0×, and 6.4× speedups over FLA v0.5.0 are reported at 256, 1.6K, and 16.4K tokens, respectively, under FP16 and CUDA Graph execution on RTX 4090.This isolated operator benchmark excludes output-epilogue fusion.Source: Figure 4 · p. 4, §2.3 · p. 5

2.1 GTR BACKBONE

  • GTR’s recurrent update and channel operations scale linearly with token count, while the recurrent state remains constant-size with respect to sequence length.Token activations and intermediate workspaces still grow linearly with sequence length.Source: §2.1 · p. 4

2.3 HARDWARE-AWARE CHUNKWISE INFERENCE

  • The inference-only CUDA operator computes chunk summaries in parallel, propagates boundary states through a lightweight recurrent scan, then computes chunk outputs independently.It is specialized for FP16 inputs and the GTR configuration; training uses reference FLA kernels.Source: §2.3 · p. 5

3 EXPERIMENTS

The experiments evaluate GTR on object detection, five additional dense-prediction tasks, and ablations of its architecture and training. They also assess inference efficiency on GPU and automotive hardware.

3.1 REAL-TIME OBJECT DETECTION

  • 58.9 AP at 1.908 ms median latency: with Objects365 pre-training, GTR-L outperforms all L-scale baselines in both metrics on COCO val2017.Latency uses compiled FP16, batch-one inference on an RTX 4090.Source: Table 1 · p. 6, §3.1 · p. 6

3.2 TRANSFER TO OTHER DENSE-PREDICTION TASKS

  • GTR achieves 45.0/49.8 mask AP for instance segmentation and 70.1/75.5 keypoint AP for pose estimation at S/X scales.GTR-X has 25.3% lower latency than RF-DETR-Seg-X; pose results improve over ECPose at both scales while reducing latency.Source: §3.2 · p. 6, Table 2 · p. 7
  • GTR-X reaches 81.3 AP50 on DOTA-v1.0; GTR-S/X reach 81.5/83.6 mIoU on Cityscapes, and depth results improve over evaluated YOLO26 variants at both scales.On DOTA-v1.0, GTR-X has lower median latency than YOLO26x-obb; on Cityscapes, GTR-X matches 83.6 mIoU with 19.5% lower latency than YOLO26-sem. NYU depth evaluation uses multi-view test-time augmentation and per-image log-affine alignment to ground-truth depth.Source: §3.2 · pp. 7–8

3.3 ABLATION STUDIES AND DISCUSSION

  • 50.7 AP with final-output alignment exceeds ViT-Linearizer at 50.0 and ViT-AdaLA at 48.2, while training from scratch reaches 31.1 AP.The comparison uses the same teacher, GTR-S student, and COCO training schedule; the ViT-Linearizer mask ratio is 0.5.Source: §3.3 · p. 8
  • Replacing learned positional embeddings with S-SwiGLU improves GTR-S by 1.2/1.3 AP at 768^2/1024^2; four-direction scanning reaches 55.5 AP on GTR-L.The GTR-L alternatives reach 55.1/55.0 AP with horizontal/vertical bidirectional scans and 54.4 AP with a single direction; replacing S-SwiGLU with learned positional embeddings under single-direction scanning yields 45.6 AP.Source: §3.3 · p. 8, Table 4 · p. 8
  • Replacing GLA in three of twelve GTR-L blocks with softmax attention yields 55.4–55.6 AP versus 55.5 AP, while raising latency from 1.908 to 2.094–2.117 ms and computation from 106 to 119 GFLOPs.All variants have 37.2M parameters and 0.19 GB peak memory.Source: §3.3 · p. 9, Table 5 · p. 9
  • Increasing GTR-S resolution from 640^2 to 1280^2 raises AP from 53.6 to 57.0 and small-object AP from 36.4 to 42.4, while latency grows 2.06× for 4× more image tokens.Median latency rises from 1.225 to 2.529 ms and peak memory from 0.096 to 0.249 GB.Source: §3.3 · p. 9, Table 6 · p. 9
  • TensorRT deployment on DRIVE AGX Thor combines standard operators with fused CUDA kernels for GLA and S-SwiGLU, enabling low-latency inference across six tasks.Across 48 task–scale–batch configurations, cosine similarity to the PyTorch FP32 reference is at least 0.9989.Source: §3.3 · p. 9

4 CONCLUSION

The paper concludes that GTR offers an accuracy–efficiency trade-off across dense-prediction tasks and demonstrates deployment on automotive hardware. The related work contrasts GTR’s recurrence and distillation with prior approaches.

4 CONCLUSION

  • GTR combines a homogeneous GLA architecture, final-output distillation, and optimized inference, with reported results supporting it as an efficient alternative to softmax attention for dense visual prediction.The conclusion also points to deployment on NVIDIA DRIVE AGX Thor as evidence of practicality on automotive hardware.Source: §4 · p. 9

AI USE STATEMENT

  • The authors report using generative AI for literature review, manuscript organization, editing, synthesis of reported results, and limited code work, with human review and verification.They state that AI tools did not generate, modify, or fabricate experimental data.Source: AI USE STATEMENT · p. 10

A RELATED WORK

B GTR ARCHITECTURE DETAILS

The appendix reports complete comparisons across six dense-prediction tasks, alongside a compact multi-scale detector encoder and query-based decoder.

B.1 MULTI-SCALE FEATURE ENCODER

  • The encoder resizes selected features from blocks 4, 8, and 12, concatenates and fuses them into stride-8/16/32 pyramid levels, then combines those levels as decoder memory.Identical source and target strides require no resampling; other features use bilinear interpolation, followed by a lightweight C2f fusion module.Source: §B.1 · pp. 17–18

B.2 QUERY-BASED DETECTION DECODER

  • The DETR-style decoder selects top-k high-classification-response anchor proposals to initialize object queries and trains with three query groups sharing decoder parameters.Only the primary query group is used at inference.Source: §B.2 · p. 18

C.1 FULL COCO OBJECT DETECTION COMPARISON

  • 53.6/57.3/58.9/59.4 AP: Objects365-pretrained GTR-S/M/L/X achieve the highest overall AP in all four COCO scale groups; GTR-L/X also have the lowest median latency in their groups.Without Objects365 pre-training, the same scale-specific COCO configurations achieve 50.7/54.0/55.5/56.2 AP.Source: §C.1 · p. 18, Table 7 · p. 19

C.2 FULL COCO INSTANCE SEGMENTATION COMPARISON

  • GTR has the highest overall mask AP in every reported COCO scale group; GTR-X improves both mask AP and latency over RF-DETR-Seg-X without SAM2 pseudo-mask supervision.Table 8 compares real-time instance-segmentation models on COCO val2017; † marks Objects365 pre-training with SAM2 pseudo masks.Source: Table 8 · p. 20, §C.2 · p. 20

C.3 FULL COCO HUMAN POSE ESTIMATION COMPARISON

  • Across the broader COCO pose-estimation comparison, GTR attains the highest keypoint AP and AR at every scale.The comparison adds RTMO, YOLO11-Pose, and ED-Pose where corresponding model scales are available.Source: §C.3 · p. 20, Table 9 · p. 21

C.4 FULL DOTA-V1.0 ORIENTED OBJECT DETECTION COMPARISON

  • Compared with RTMDet-R-l, GTR-X gains 0.8 AP50 with 6.0M fewer parameters; versus Oriented-DETR (Swin-T), it gains 1.5 AP50 and is 3.9× faster.GTR-X also has the highest reported plane and harbor AP50, at 89.5 and 84.5.Source: §C.4 · p. 21, Table 10 · p. 22

C.5 FULL CITYSCAPES SEMANTIC SEGMENTATION COMPARISON

  • GTR-S/M/L exceed size-matched YOLO26-sem by 0.7/1.0/0.3 mIoU points, while GTR-X matches YOLO26x-sem at 83.6 mIoU with fewer parameters, FLOPs, and median latency.The GTR-X reductions are 20.0% fewer parameters, 63.2% fewer FLOPs, and 19.5% lower median latency.Source: Table 11 · p. 22, §C.5 · p. 22

C.6 FULL NYU DEPTH ESTIMATION COMPARISON

  • At every evaluated scale, GTR improves δ1, AbsRel, and RMSE over the corresponding YOLO26 depth variant while using fewer GFLOPs; GTR-X reduces median latency from 2.730 to 2.158 ms.ZipDepth is re-evaluated on the same filled ground-truth data as YOLO26 and GTR; its original paper reports metrics under its own protocol.Source: Table 12 · p. 23, §C.6 · p. 23

C.7 DETAILED ABLATION RESULTS

The ablations examine spatial modeling, scan direction, hybrid softmax attention, input resolution, and Objects365 detector pre-training.

Spatial mechanism

  • S-SwiGLU improves overall AP over learned positional embeddings at both tested resolutions, with only a small FLOP increase.Table 13 compares the spatial mechanisms on GTR-S; S-SwiGLU uses no explicit positional embedding.Source: §C.7 · p. 23, Table 13 · p. 23

Scan schedule

  • Four-direction scanning yields the highest overall AP among evaluated GTR-L scan schedules; with single-direction scanning, removing S-SwiGLU changes AP more than reducing scan directions.Table 14 compares scan direction and spatial mechanism; LR/RL/TB/BT denote horizontal and vertical scan directions.Source: §C.7 · p. 23, Table 14 · p. 23

Hybrid attention

  • Replacing GLA with softmax attention in three blocks raises GFLOPs from 106 to 119 and median latency by 9.7–11.0%, while overall AP stays within 0.1 of pure GLA.The hybrids use FlashAttention-2; reported memory values coincide at displayed precision.Source: §C.7 · p. 24

Resolution scaling

  • Most of the accuracy gain in the GTR-S resolution sweep occurs for small and medium objects; batch-one rate is computed as the reciprocal of median latency.The sweep uses Objects365-pre-trained detector initialization.Source: §C.7 · p. 24

Detector pre-training

  • Objects365 detector pre-training raises GTR-S overall AP by 2.9 and small-object AP by 5.3.Table 15 reports the metric breakdown with and without Objects365 pre-training; the four-scale results are in Table 7.Source: §C.7 · p. 24, Table 15 · p. 24

D EXPERIMENTAL DETAILS

The experimental details specify evaluation and training protocols, model configurations, and task-specific fine-tuning for the reported comparisons.

D EXPERIMENTAL DETAILS

  • The detector comparison evaluates COCO val2017 AP and related metrics after COCO fine-tuning, and reports latency as median per-image time under a unified FP16 batch-one RTX 4090 protocol.Timing excludes compilation, graph capture, autotuning, and warm-up; DETR-style measurements cover the network forward pass, while YOLO variants include CUDA NMS when required.Source: D EXPERIMENTAL DETAILS · p. 24

D.1.1 ARCHITECTURE CONFIGURATIONS

  • All GTR variants use a 12-layer, stride-16 GLA backbone with intermediate multi-scale features; GTR-X keeps GTR-L token-mixing width but expands SwiGLU and decoder FFN dimensions.From GTR-S to GTR-L, embedding dimension and GLA heads increase from 192/3 to 384/6; GTR-X increases the SwiGLU expansion ratio from 4 to 6 and decoder FFN dimension from 1,024 to 2,048.Source: §D.1.1 · p. 24, Table 16 · p. 25

D.1.2 FINAL-OUTPUT REPRESENTATION ALIGNMENT

  • Final-output alignment uses a frozen detection-specialized teacher to provide patch-token targets, with one representation-distillation stage before downstream supervised training.The image pool combines ImageNet-1K and COCO train2017; all variants train for 300 epochs, with longer warm-up and higher drop-path for GTR-L/X.Source: §D.1.2 · p. 25, Figure 6 · p. 26

D.1.3 OBJECTS365 PRE-TRAINING

  • All detector scales receive the same 36-epoch Objects365 pre-training schedule, while backbone learning rates decrease with increasing model capacity.The schedule uses global batch size 128 and base learning rate 2 × 10^-3; backbone rates are 10^-4 for GTR-S/M, 2 × 10^-5 for GTR-L, and 10^-5 for GTR-X.Source: §D.1.3 · pp. 25–26, Table 18 · p. 26

D.1.4 COCO FINE-TUNING

  • After Objects365 pre-training, COCO fine-tuning loads the detector checkpoint except for the classification head and uses a fixed 30-epoch schedule across scales.All variants train on COCO train2017 and are evaluated on val2017.Source: §D.1.4 · p. 26, Table 19 · p. 26

D.1.5 HUMAN POSE ESTIMATION

  • For pose estimation, only the GTR backbone is initialized from the COCO detection checkpoint; all pose-specific parameters are trained from scratch.All variants use AdamW, a base learning rate of 5 × 10^-4, and global batch size 32.Source: §D.1.5 · p. 26, Table 20 · p. 27

D.1.6 SEMANTIC SEGMENTATION

  • For Cityscapes semantic segmentation, task-specific parameters are trained from scratch, and evaluation uses sliding windows with 1024×1024 windows and 768-pixel stride.Training uses fine annotations only, random resizing, horizontal flipping, and 1024×1024 random crops; the model retains the scale=2.0 neck branch and uses a lightweight FCN head.Source: §D.1.6 · p. 27, Table 21 · p. 27

D.1.7 MONOCULAR DEPTH ESTIMATION WITH DIVERSE TRAINING DATA

The appendix specifies mixed-source depth training and evaluation, oriented-detection adaptation, and TensorRT deployment and timing protocols. On DRIVE AGX Thor, batch-one latency spans 2.282–8.769 ms across tasks and scales, with roofline analysis indicating memory-traffic limits.

D.1.7 MONOCULAR DEPTH ESTIMATION WITH DIVERSE TRAINING DATA

  • Depth training uses an eight-source measured and pseudo-labeled mixture, with nested Depth Anything 3 generating metric-depth pseudo-labels without additional scale calibration.The protocol oversamples measured datasets and retains metric-depth-preserving augmentations; composition-based geometric augmentations are disabled.Source: §D.1.7 · pp. 28–29
  • NYU Depth V2 evaluation uses 654 images without NYU-specific fine-tuning, with six test-time views averaged before per-image log-affine ground-truth alignment.The models initialize the backbone from a pretrained checkpoint and train task-specific parameters from scratch; the depth head uses SiLog loss and clamped log-depth decoding.Source: §D.1.7 · pp. 28–29

D.1.8 ORIENTED OBJECT DETECTION

  • DOTA-v1.0 training and testing use three image scales, overlapping 1024 × 1024 crops, and an oriented-box decoder with angle distribution refinement and rotated deformable-attention offsets.The decoder also uses oriented contrastive denoising; regression combines ℓ1 supervision with Gaussian KLD loss.Source: §D.1.8 · p. 30, Table 23 · p. 31

E.2.1 HARDWARE, SOFTWARE, AND DEPLOYMENT

  • Thor-targeted operators are compiled for sm 110 and integrated into TensorRT, and all reported deployment results use FP16.These are distinct builds from the sm 89 PyTorch-extension GLA operator benchmark.Source: §E.1 · p. 31, Table 24 · p. 32

E.1 DRIVE AGX THOR DEPLOYMENT

  • Across six tasks and four scales, optimized Thor operating points move upward and rightward from the TensorRT baseline by eliminating intermediate memory traffic.The plotted points lie near bandwidth roofs and left of the 1,032 FLOP/byte ridge point.Source: §E.1 · p. 31, Figure 7 · p. 32
  • 2.282–8.769 ms batch-one latency spans tasks and model scales on DRIVE AGX Thor, while optimized roofline points indicate memory-traffic limits.The optimized points reach 89–141% of the nominal DRAM roof; batch-one inference also gives higher throughput than batch eight for 1024 × 1024 oriented detection and semantic segmentation.Source: §E.1 · p. 31, Figure 7 · p. 32, Table 25 · p. 33

E.2.2 TIMED SCOPE AND POST-PROCESSING

  • Latency comparisons include forward pass plus CUDA NMS for YOLOv9, YOLO11, and YOLOv12, but only forward pass for DETR-style models.NMS-free detector evaluation uses a fixed 1,000-image COCO subset, whereas NMS-dependent models use all 5,000 val2017 images.Source: §E.2.2 · pp. 33–34

E.2.3 WARM-UP AND STATISTICS

  • The RTX 4090 batch-one protocol uses CUDA-event timing, and median per-image latency is the comparison statistic.Each model is benchmarked in an independent process; reported statistics also include minimum and arithmetic mean latency.Source: §E.2.1 · pp. 31–32, §E.2.3 · p. 34

E.3.1 IMPLEMENTATION DETAILS

The specialized GLA operator targets batch-one FP16 GTR inference by parallelizing chunk summaries while retaining a lightweight sequential boundary-state scan. The appendix also presents qualitative detection examples under challenging image conditions.

E.3.1 IMPLEMENTATION DETAILS

  • The custom CUDA operator targets batch-one FP16 inference with fixed chunk and head dimensions, while preserving the recurrence and changing only its execution schedule and fusion.It uses a zero initial state and omits backward propagation and final-state output supported by the general FLA implementation.Source: §E.3.1 · p. 34
  • Parallel chunk summaries followed by a coalesced linear boundary-state scan retain O(L) work for fixed head dimensions and chunk size.The authors report this scan is faster in the target regime than a tree-based prefix scan, which adds global-memory passes and strided accesses.Source: §E.3.1 · p. 34
  • The implementation fuses chunk-local sums and contractions, avoids materializing intra-chunk scores globally, and uses length-aware dispatch and an optional fused scan-output path.The end-to-end epilogue also fuses RMSNorm and the output gate, removing a kernel launch and intermediate memory round trip; that epilogue fusion is excluded from isolated operator measurements.Source: §E.3.1 · pp. 34–35

F.1 OBJECT DETECTION RESULTS

  • GTR-X detection examples show localization of small, overlapping instances under crowding, occlusion, motion blur, object rotation, and low illumination.The qualitative appendix also presents transfer examples spanning six dense-prediction tasks and datasets.Source: §F.1 · p. 35, Figure 8 · p. 35, F QUALITATIVE RESULTS · p. 35

F.2 INSTANCE SEGMENTATION RESULTS

The qualitative instance-segmentation examples show GTR-X masks following visible contours and separating adjacent instances on the challenging images also used for detection.

F.2 INSTANCE SEGMENTATION RESULTS

  • GTR-X masks follow visible object contours and separate adjacent instances on the same challenging images shown for object detection.Figure 9 visualizes the selected predictions; the text connects them with Table 8 as an illustration of transfer to instance segmentation.Source: §F.2 · p. 35, Figure 9 · p. 36

F.3 HUMAN POSE ESTIMATION RESULTS

The qualitative results show GTR-X pose predictions for overlapping people in difficult scenes and compare semantic predictions with Cityscapes annotations. GTR-L is also applied to indoor and outdoor depth imagery without dataset-specific fine-tuning.

F.3 HUMAN POSE ESTIMATION RESULTS

  • Figure 10 shows GTR-X predicting articulated poses for multiple overlapping people under crowding, low illumination, and motion blur.The examples are qualitative human-pose results.Source: §F.3 · p. 36

F.4 SEMANTIC SEGMENTATION RESULTS

  • Figure 11 compares GTR-X semantic predictions with Cityscapes ground truth, with each row showing the RGB input, prediction, and annotation.The visual results are qualitative semantic-segmentation examples.Source: §F.4 · p. 37, Figure 11 · p. 37

F.5 ZERO-SHOT MONOCULAR DEPTH ESTIMATION RESULTS

  • GTR-L is applied without dataset-specific fine-tuning to indoor DTU and outdoor nuScenes imagery for zero-shot monocular depth estimation.Source: §F.5 · p. 37

F.5.1 INDOOR DTU RESULTS

The DTU depth examples use projected calibrated point-cloud geometry as reference and correct each predicted view’s global scale before multi-view fusion. These visualizations therefore show geometry after reference-assisted scale correction, not unadjusted zero-shot scale.

F.5.1 INDOOR DTU RESULTS

  • Reference depth is projected from calibrated DTU point-cloud geometry, covers 67.9% of pixels on average, and lacks support mainly on background fabric and specular surfaces.Figure 12 compares RGB inputs, GTR-L predictions, and projected references.Source: §F.5.1 · p. 37, Figure 12 · p. 38
  • Each DTU prediction is rescaled by the median reference-to-prediction depth ratio over valid pixels, correcting global scale while preserving within-view relative depth.The scale correction addresses view-dependent scale drift before fusion.Source: §F.5.1 · p. 37, Figure 13 · p. 38
  • The rescaled depths from all 49 calibrated views are back-projected and fused into a TSDF volume with 2-mm voxels to render colored points.The reconstructions use reference-assisted scale correction.Source: §F.5.1 · p. 37, Figure 13 · p. 38
  • Figure 14 shows outdoor nuScenes RGB inputs and their corresponding GTR-L depth predictions in matching row-major order.The caption identifies six surround-view inputs followed by predictions.Source: Figure 14 · p. 39

F.5.2 OUTDOOR NUSCENES RESULTS

GTR-L produces surround-view depth predictions across nuScenes scenes without nuScenes-specific fine-tuning, and the paper visualizes temporally fused maps after LiDAR-assisted scale correction. GTR-X’s DOTA-v1.0 examples show oriented boxes tracking objects at varied angles, including crowded scenes.

F.5.2 OUTDOOR NUSCENES RESULTS

  • GTR-L predicts depth for six synchronized nuScenes cameras across daytime, nighttime, and roadside-construction scenes without nuScenes-specific fine-tuning.Figures 14–16 preserve camera ordering between RGB inputs and corresponding depth predictions.Source: §F.5.2 · p. 39, Figure 15 · p. 40, Figure 16 · p. 40
  • Local nuScenes maps aggregate frames within 4 seconds before and after each reference time, with LiDAR-assisted per-frame scale alignment before TSDF fusion.The resulting examples qualitatively illustrate geometric consistency after scale correction.Source: §F.5.2 · p. 39, Figure 17 · p. 41, Figure 18 · p. 41

F.6 ORIENTED OBJECT DETECTION RESULTS

  • On DOTA-v1.0 test images, GTR-X’s oriented boxes track ships and vehicles at varied angles and remain separated in densely packed harbor and road scenes.Figure 19 provides qualitative examples complementing Table 10’s quantitative results.Source: §F.6 · p. 39, Figure 19 · p. 42
Loading 2609.26590v2…