Autonomous highway driving, especially for long-haul heavy trucks, requires detecting objects at long ranges beyond 500 meters to satisfy braking distance requirements at high speeds. At long distances, vehicles and other critical objects occupy only a few pixels in high-resolution images, causing state-of-the-art object detectors to fail. This challenge is compounded by the limited effective range of commercially available LiDAR sensors, which fall short of ultra-long range thresholds because of quadratic loss of resolution with distance, making image-based detection the most practically scalable solution given commercially available sensor constraints. We introduce Telescope, a two-stage detection model designed for ultra-long range autonomous driving. Alongside a powerful detection backbone, this model contains a novel re-sampling layer and image transformation to address the fundamental challenges of detecting small, distant objects. Telescope achieves 76% relative improvement in mAP in ultra-long range detection compared to state-of-the-art methods (improving from an absolute mAP of 0.185 to 0.326 at distances beyond 250 meters), requires minimal computational overhead, and maintains strong performance across all detection ranges.
Figures & tables
Figure 1: Long-range Objects in Driving Datasets. Analysis of the TruckDrive [ 16 ] dataset shows the distribution of object distances and the breakdown of the pixel-wise composition of objects at each distance. While all object ranges are equally represented in images, the proportion of pixel area disproportionately favors nearby objects, with long ( 150−250 m) and ultra-long ( ≥250 m) objects occupying only a small fraction of image pixels.
Figure 2: Telescope. We propose a two-stage ultra-long range detection model. Stage one uses a down-sampled image to estimate the hyperbolic foveation image transformation parameters. This transformation enlarges distant objects at the center of the transform while shrinking nearby objects at the periphery. Stage two uses this transformed image alongside learned hyperbolic embeddings to detect objects at distances of up to 1km.
Figure 3: Hyperbolic Foveated Transform. The transformation coefficients enable the re-parameterization of the bounding box in the induced Riemannian space where the box center and tangent vector magnitudes fully describe the box location and shape.
Figure 4: gIoU for Multi-Scale Training. The gIoU metric is a better training loss for multi-scale objects as it provides a gradient even when bounding boxes are non-overlapping.
Figure 5: Learned hyperbolic foveated transform on the TruckDrive dataset. The original image (left) and the foveated image (right) are shown, together with the percentage increase in object bounding-box area. Both views are cropped to the same image region, highlighting the local magnification induced by the foveation transform. The proposed transform is effective for both isolated targets and dense, busy scenes.
Method
COCO
mAP
mAP 0-50
mAP 50-150
mAP 150-250
mAP 250+
Deformable DETR
0.32
0.52
0.34
0.24
0.17
+ De-noising
0.50 ( +0.18 )
0.69 ( +0.17 )
0.48 ( +0.14 )
0.32 ( +0.08 )
0.29 ( +0.12 )
+ Hyperbolic Foveation
0.50 ( +0.00 )
0.61 ( -0.08 )
0.50 ( +0.02 )
0.34 ( +0.02 )
0.33 ( +0.03 )
Table 1: Ablation Experiments on the Components of Telescope . A de-noising training scheme helps improve model performance and the hyperbolic foveation improves mAP at far, long, and ultra-long ranges while slightly reducing performance for nearby objects. All rows use SAM3 encoder.
Figure 6: Qualitative Visualization. Detections from Telescope on the TruckDrive [ 16 ] dataset. Telescope consistently detects and localizes distant vehicles that occupy only a few pixels, while preserving accurate predictions for nearby objects. These examples highlight the effect of the proposed hyperbolic foveated transform in magnifying ultra-long range regions and improving sensitivity to objects in these ranges. Cut-outs provide a high-resolution, zoomed-in region to more clearly visualize objects.
Method
COCO
PASCAL
mAP
mAP 0-50
mAP 50-150
mAP 150-250
mAP 250+
mAP 50
mAP 75
DINOv2 + Deformable DETR
0.186
0.432
0.208
0.108
0.042
0.375
0.161
DINOv3 + Deformable DETR
0.212
0.467
0.250
0.137
0.059
0.419
0.190
SAM3 + Deformable DETR
0.317
0.523
0.344
0.244
0.171
0.521
0.329
SAM3 + Denoising [ 56 ]
0.501
0.692
0.483
0.321
0.292
0.758
0.545
Table 2: Distance-wise Ablation Experiments of Encoder and Head . All models perform well for objects close to the ego vehicle, but have degraded performance as distance increases. The proposed model using the SAM3 image encoder backbone, Deformable DETR detection head, and denoising-based training to provide optimal performance, especially at ultra-long range distances and across all classes.
Method
COCO mAP
Person
Bike
Sign
Car
Truck
Debris
DINOv2 + Deformable DETR
0.154
0.302
0.175
0.302
0.301
0.092
DINOv3 + Deformable DETR
0.149
0.304
0.222
0.345
0.334
0.120
SAM3 + Deformable DETR
0.106
0.236
0.415
0.459
0.442
0.245
SAM3 + Denoising [ 56 ]
0.330
0.436
0.617
0.631
0.596
0.451
Table 3: Class-wise Ablation Experiments for Backbone and Head . Results for the TruckDrive dataset for each class for existing state-of-the-art models as well as foundation model ablations. Leveraging foundation models as pre-trained image encoders provides a strong prior for object detection comparable to state-of-the-art specialized models. The proposed model using the SAM3 image encoder backbone, Deformable DETR detection head, and de-noising training scheme to provide optimal performance across all classes.
Figure 7: Qualitative Comparison. Qualitative comparison between the proposed method, Telescope, and state-of-the-art baselines. Both RVSA [ 44 ] and RFLA [ 52 ] are specialized for small object detection while DETR [ 5 ] and DINO [ 56 ] are strong general object detectors, but perform worse in long an ultra-long range object detection. Ground truth annotations are shown on the left. Zoomed-in views corresponding to the red rectangles are provided to highlight detections at long and ultra-long range, where some objects reach up to 1 km. All baselines are fine-tuned on the TruckDrive [ 16 ] dataset.
Method
COCO
PASCAL
mAP
mAP 0-50
mAP 50-150
mAP 150-250
mAP 250+
mAP 50
mAP 75
DETR [ 5 ]
0.166
0.396
0.178
0.081
0.072
0.335
0.147
Grounding DINO [ 28 ]
0.286
0.376
0.262
0.147
0.156
0.417
0.296
FOVEA [ 42 ]
0.113
0.169
0.086
0.008
0.005
0.189
0.115
YOLO11x [ 19 ]
0.266
0.421
0.218
0.134
0.117
0.510
0.195
DINO [ 56 ]
0.222
0.335
0.239
0.189
0.179
0.371
0.226
Table 4: Distance-Wise Ultra-long Range Object Detection Evaluation on TruckDrive dataset . The proposed model using the SAM3 image encoder backbone, Deformable DETR detection head, de-noising training, and TeleScope re-sampling layer provides optimal performance, especially at ultra long range distances, where it significantly improves over previous methods.
Method
COCO mAP
Person
Bike
Sign
Car
Truck
Debris
DETR [ 5 ]
0.222
0.327
0.179
0.299
0.247
0.083
Grounding DINO [ 28 ]
0.141
0.174
0.472
0.591
0.317
0.024
FOVEA [ 42 ]
0.028
0.043
0.060
0.377
0.056
0.107
YOLO11x [ 19 ]
0.172
0.400
0.370
0.330
0.376
0.049
DINO [ 56 ]
0.059
0.819
0.264
0.431
0.334
0.226
Table 5: Class-Wise Ultra-Long Range Detection Evaluation on TruckDrive dataset . Leveraging foundation models as pre-trained image encoders provides a strong prior for object detection comparable to state-of-the-art specialized models. The proposed model using the SAM3 image encoder backbone, Deformable DETR detection head, de-noising training, and TeleScope re-sampling layer provides optimal performance across all classes.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Long-range Objects in Argoverse Driving. Analysis of the Argoverse [ 47 ] dataset shows the distribution of object distances and the breakdown of the pixel-wise composition of objects at each distance. Multiple object ranges are represented in images, but nearby objects are disproportionately favored in terms of pixel area, with far (50-150m) and long (150-250m) range objects occupying only a small fraction of image pixels.
Class
Person
Bike
Car
Sign
Truck
Debris
Average Height [m]
0.70
0.71
1.89
1.26
2.90
0.41
Appendix
Table 6: Average Class Heights . Average heights across the class used to compute the approximate object distance given bounding box height and camera focal length.
Method
COCO
PASCAL
mAP
mAP 0-50
mAP 50-150
mAP 150-250
mAP 50
mAP 75
UniverseNet [ 38 ]
0.123
0.156
0.042
0.016
0.271
0.100
RVSA [ 44 ]
0.121
0.150
0.052
0.026
0.260
0.098
RFLA [ 52 ]
0.106
0.131
0.036
0.023
0.250
0.070
Telescope (Ours)
0.232
0.268
0.104
0.036
0.502
0.177
Appendix
Table 7: Distance-Wise Object Detection Evaluation on Argoverse 2 Dataset . The proposed model using the SAM3 image encoder backbone, Deformable DETR detection head, de-noising training, and TeleScope re-sampling layer provides optimal performance across all distances.
Method
COCO mAP
Regular
Pedest.
Bollard
Const.
Stop
Bicycle
Wheeled
Vehicle
Barrel
Sign
Device
UniverseNet [ 38 ]
0.396
0.251
0.069
0.353
0.209
0.151
0.098
RVSA [ 44 ]
0.381
0.183
0.099
0.192
0.120
0.151
0.091
RFLA [ 52 ]
0.352
0.233
0.044
0.269
0.153
0.124
0.070
Telescope (Ours)
0.562
0.412
0.154
0.417
0.316
0.355
0.259
Appendix
Table 8: Class-Wise Object Detection Evaluation on Argoverse 2 Dataset . Leveraging foundation models as pre-trained image encoders provides a strong prior for object detection comparable to state-of-the-art specialized models. The proposed model using the SAM3 image encoder backbone, Deformable DETR detection head, de-noising training, and TeleScope re-sampling layer provides optimal performance across all classes. Shown are results for the 7 most common classes in Argoverse 2.
Figure 9: Additional Qualitative Comparison on Argoverse Dataset. Qualitative comparison between the proposed method, Telescope, and state-of-the-art baselines specialized for small object detection. Ground truth annotations are shown on the left. Notably, there are many target boxes which represent occluded objects (rows 1, 2, 3, 6, and 7). All methods are fine-tuned on the Argoverse [ 47 ] dataset.
Figure 10: Additional Qualitative Comparison on TruckDrive Dataset. Qualitative comparison between the proposed method, Telescope, and state-of-the-art baselines. Both RVSA [ 44 ] and RFLA [ 52 ] are specialized for small object detection while DETR [ 5 ] and DINO [ 56 ] are strong general object detectors, but perform worse in long an ultra-long range object detection. Ground truth annotations are shown on the left. Zoomed-in views corresponding to the red rectangles are provided to highlight detections at long and ultra-long range, where some objects reach up to 1 km. All baselines are fine-tuned on the TruckDrive [ 16 ] dataset.
Component
Sub-Component
Layer
Parameters
SAM3 Backbone
ViT (ViTDet)
Patch Embed
patch: 14, dim: 1024
Transformer
depth: 32, heads: 16, mlp: 4.625 ×
FPN Neck
ConvTranspose2d
× 2 upsample per scale
Conv2d
(1024, 256), k=1; (256, 256), k=3
Foveation Estimation
Head
AdaptiveAvgPool2d
output: 1 × 1
MLP
(256, 128, 4), ReLU, Sigmoid/Softplus
Appendix
Table 9: Network Details of SAM3 + DINO 2-Stage Model.
Safe highway autonomy for heavy trucks remains an open and unsolved challenge: due to long braking distances, scene understanding of hundreds of meters is required for anticipatory planning and to allow safe braking margins. However, existing driving datasets primarily cover urban scenes, with perception effectively limited to short ranges of only up to 100 meters. To address this gap, we introduce TruckDrive, a highway-scale multimodal driving dataset, captured with a sensor suite purpose-built for long range sensing: seven long-range FMCW LiDARs measuring range and radial velocity, three high-resolution short-range LiDARs, eleven 8MP surround cameras with varying focal lengths and ten 4D FMCW radars. The dataset offers 475 thousands samples with 165 thousands densely annotated frames for driving perception benchmarking up to 1,000 meters for 2D detection and 400 meters for 3D detection, depth estimation, tracking, planning and end to end driving over 20 seconds sequences at highway speeds. We find that state-of-the-art autonomous driving models do not generalize to ranges beyond 150 meters, with drops between 31% and 99% in 3D perception tasks, exposing a systematic long-range gap that current architectures and training signals cannot close.
Filippo Ghilotti, Edoardo Palladin, Samuel Brucker +3
3D object detection is the backbone of perception for automated vehicles (AV) and broader intelligent transportation systems applications. Long-range detection is challenging because sensing evidence is sparse; yet this long-range'' scenario is routine in traffic. Although >30m is often labeled long-range in computer vision, on roadways it affords only approx. 1-2s for perception and decision-making. Under such extreme sparsity, two core challenges arise. First, early multimodal fusion tends to discard sparsity information and inject noise from empty or falsely occupied cells, degrading long-range recall. Second, context-agnostic uniform channel supervision favors dense and near-range samples, leaving far and small objects under-optimized, delaying the earliest detection of distant objects. We propose Ask The Neighbor'' (ATN3D), a LiDAR-Radar framework tailored for sparse-range conditions. ATN3D introduces (i) Density-aware early fusion with cross-modal gating that conditions fusion on per-voxel density/sparsity and Radar evidence, (ii) Occupancy-gated neighborhood aggregation with circular kernels to aggregate only from credible cells, (iii) Evidence-conditioned channel self-attention to adapt channel weights with weather/range, and (iv) a Range-aware loss that re-balances classification and localization by distance, aligning training with distance-stratified evaluation. On the VoD benchmark across clear and foggy conditions, ATN3D surpasses strong baselines: +3.55% mAP in clear weather and +8.41% mAP under simulated heavy fog; for >30m objects, gains are +3.33% (clear) and +2.09% (heavy fog). These results indicate earlier and more reliable long-range detections under sparse sensing in on-road traffic.
Long-range 3D object detection is critical for safe autonomous driving at highway speeds, yet existing radar-camera fusion methods remain limited at extended ranges. BEV-based methods capture scene-level context but incur rapidly growing computation and often lose fine-grained object detail, while query-based methods are efficient but provide limited scene-level context. Temporal fusion further requires both multi-frame accumulation for sparse distant observations and object-level motion modeling for fast-moving objects. We propose Horizon3D, a sparse radar-camera fusion framework for long-range 3D object detection that combines Gaussian primitives with sparse BEV features. Horizon3D initializes Gaussian primitives at radar- and camera-estimated object keypoints using Keypoint-Guided Gaussian Initialization, refines them through Object-Centric Sparse Fusion, and splats them onto the BEV plane to fuse object-level detail with sparse radar BEV context. It further introduces Dual-Path Temporal Fusion, which aggregates temporal cues through a BEV path for scene-level accumulation and a Gaussian path for object-level motion propagation. Experiments on TruckScenes show that Horizon3D achieves state-of-the-art radar-camera 3D detection performance. On the validation set, it outperforms the previous best method by +3.0 NDS and +1.6 mAP while maintaining competitive inference speed.
Geonho Bang, Geunju Baek, Dongyoung Lee +2
Seoul National University, Seoul, Republic of Korea