Far3D: Expanding the Horizon for Surround-view 3D Object Detection

• Most surround-view methods focus on close-range perception, leaving long-range detection underexplored despite its importance for maintaining safe distances.The cited work gives approximately 50 meters on nuScenes as an example of close-range focus.Source: §1 · p. 1
• Sparse global queries avoid dense computation but can miss distant targets; 2D detector recall is higher, although 2D recall is not a 3D upper bound.Directly converting 2D proposals can also create wrong-depth redundancy and range-amplified 3D deviations.Source: §1 · pp. 1–2, Figure 1 · p. 1
• Far3D combines 2D-prior-derived adaptive queries with global queries, perspective-aware feature aggregation, and range-modulated denoising.Aggregation selects informative features across views and scales, while denoising addresses query error propagation and convergence issues.Source: §1 · p. 2
• Far3D reports state-of-the-art performance on long-range Argoverse 2, surpasses prior LiDAR-based methods, and achieves state-of-the-art metrics on nuScenes.Source: §1 · p. 2

2.1 Surround-view 3D Object Detection

• BEV methods transform image features into BEV space, while sparse-query methods use lightweight global object queries to predict 3D boxes.Source: §2.1 · p. 2
• Temporal methods use prior frames in different ways; StreamPETR propagates long-term historical information through object queries, as does Far3D's temporal modeling.Source: §2.1 · p. 3, Figure 3 · p. 3

2.2 2D Auxiliary Tasks for 3D Detection

• Prior 2D auxiliary approaches include proposal generation and 2D supervision; Far3D targets long-range detection with 2D-derived adaptive queries alongside global queries.It also integrates perspective-aware aggregation and 3D denoising to address task-specific issues.Source: §2.2 · p. 3

3.1 Overview

• The decoder combines 3D adaptive and global queries, then applies query self-attention and perspective-aware aggregation between queries and image features.Adaptive queries are generated from reliable 2D proposals and depths transformed into 3D; image features are extracted at multiple levels.Source: §3.1 · p. 3

3.2 Adaptive Query Generation

• Reliable 2D boxes and corresponding depths are filtered and projected into 3D proposals, whose adaptive queries sum positional and semantic embeddings.The proposals use 2D box centers and depth; semantic context is sampled from image features and paired with the 2D box confidence.Source: §3.2 · p. 4

3.3 Perspective-aware Aggregation

• Because suitable feature levels vary with object range, perspective-aware aggregation uses 3D reference points and learned offsets to sample features across views and scales.The aggregated features contribute to 3D queries according to their relative importance.Source: §3.3 · p. 4

3.4 Range-modulated 3D Denoising

• Range-modulated denoising creates positive and negative noisy queries around ground-truth objects, with position-aware noise tied to object position and scale.Each group contains one positive and K negative samples; positive offsets are constrained within ground-truth boxes, while negative offsets vary with position range.Source: §3.4 · pp. 4–5
• On Argoverse 2 at 150 meters, Far3D outperforms prior surround-view methods and several state-of-the-art LiDAR methods; on nuScenes, it achieves the highest performance among prior arts.The Argoverse 2 evaluation covers 26 object categories.Source: Table 1 · p. 5, Table 2 · p. 5

4.1 Datasets and Metrics

• Argoverse 2 evaluation covers 26 categories over 150 meters and uses mAP, CDS, ATE, ASE, and AOE; nuScenes comparisons use mAP and NDS.Argoverse 2 provides seven high-resolution ring cameras with a combined 360° field of view.Source: §4.1 · p. 5

4.2 Implementation Details

• The main Argoverse 2 experiments use VoV-99 with 960×640 inputs, while the scaled model uses ViT-L at 1536×1536 resolution.On nuScenes, the method uses a ResNet101 backbone; the Argoverse 2 range is set to 152.4m × 152.4m.Source: §4.2 · pp. 5–6, §4.3 · p. 6

4.3 Main Results

• Far3D outperforms other methods across 0–150m and 50–150m recall and mAP comparisons, with greater gains for distant objects at 2m and 4m thresholds; it reports 51.0% mAP and 59.4% NDS on nuScenes val.On nuScenes test, it reports 63.5% mAP and 68.7% NDS.Source: §4.3 · p. 6

4.4 Ablation Study & Analysis

• Adaptive queries add 2.1% mAP and 1.5% CDS, perspective-aware aggregation adds 1.0% mAP and 1.2% CDS, and range-modulated denoising adds 1.0% mAP and 0.8% CDS over sequential ablations.The framework remains robust as the number of global queries decreases, unlike StreamPETR, which has convergence problems with few global queries.Source: Table 3 · p. 6, §4.4 · p. 7, Table 6 · p. 7

5 Conclusion

• Far3D is a sparse query-based method for long-range 3D detection, with 2D-derived adaptive queries, perspective-aware aggregation, and range-modulated denoising.The authors report that long-tail class performance lowers Argoverse 2 average precision and question unified metrics for close- and long-range objects.Source: §5 · p. 7

6 Supplementary

• For temporal modeling, Far3D propagates high-scoring queries from previous frames, selected independently of original query type, following StreamPETR.Figure 6 depicts global, adaptive, and propagated queries.Source: §6 · p. 10, Figure 6 · p. 10
• Far3D targets long-range detection with adaptive queries and denoising, whereas MV2D focuses on conventional 2D-to-3D detection and BEVFormer v2 on perspective supervision.The authors report that distant objects make convergence difficult for existing methods and that their denoising alleviates this issue.Source: §6 · p. 10
• With 150m training versus StreamPETR’s 50m, Far3D alleviates close-range degradation and improves performance by a larger scale in the 50-150m range.Table 7 compares results in 0-50m and 50-150m.Source: Table 7 · p. 10
• Far3D reaches 51.0 mAP versus MV2D’s 47.1 mAP on nuScenes val, and 63.5 mAP versus BEVFormer v2’s 55.6 mAP on test.The paper also reports exceptional performance on Argoverse 2.Source: §6 · p. 10
• Replacing an average of 92 adaptive queries with extra queries decreases performance by 1.5% mAP and 1.3% mCDS.The average adaptive-query count per Argoverse 2 sample ranges from 11 to 236.Source: §6 · p. 10
• Across 26 Argoverse 2 categories and the 0-50m and 50-100m ranges, Far3D consistently achieves the best results.Message Board Trailer is excluded from the all-category comparison because its result is near zero.Source: §6 · p. 10, Figure 8 · p. 11, Figure 9 · p. 11
• Far3D iteratively updates queries with deformable attention over FPN features p2-p5, enabling adaptive matching across object distances while reducing computational complexity.The learned feature-layer choices align closely with manually selected layers based on object distance.Source: §6 · p. 12