Generalized panorama depth estimation has witnessed great advances in recent years. These methods typically achieve robust zero-shot generalization performance using large pretrained models, but their high computational cost makes real-time inference on embedded platforms challenging. In contrast, previous lightweight panorama models have mostly been developed for single-domain evaluations and cannot generalize well due to domain shifts. To reduce the gap between these two types of models, we introduce EG360, an efficient and generalizable panorama depth estimation model. Specifically, building on a lightweight depth model, we design an Adaptive Panorama-Perspective Distillation (APPD) strategy guided by scene geometry, along with a Spherical Cross-Level Feature Fusion (SCFF) module for global semantic perception. As a result, with only 6.593M parameters after reparameterization, EG360 achieves strong zero-shot performance on real-world benchmarks and runs at 35 FPS on a Jetson Orin NX 16GB in MAXN mode, whereas a previous ViT-small model runs at less than 2 FPS. Our code will also be made public for future research.
Figures & tables
Figure 1: Accuracy–computation trade-off for panoramic depth estimation. Mean AbsRel is averaged equally over Matterport3D, Stanford2D3D, and Metropolis. Filled marker area is positively proportional to measured inference FPS on a Jetson Orin NX 16GB (MAXN mode, PyTorch eager, AMP FP16, batch size 1). EG360 uses reparameterization for inference. PaGeR ran into out-of-memory (OOM) and is represented by a fixed-size hollow diamond.
Datasets
Indoor
Outdoor
Samples
Labeled Synthetic Datasets
Structured3D [ 59 ]
✓
18,222
PanoCARLA [ 36 ]
✓
112,212
TartanAirV2 [ 48 ]
✓
✓
432,381
Labeled total
562,815
Unlabeled Real-world Datasets
Table 1: Labeled synthetic and unlabeled real-world datasets used.
Stanford2D3D
Metropolis
Teacher
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
DAP [ 23 ]
0.071
0.965
0.185
0.733
MoGe2 [ 47 ]
0.150
0.774
0.127
0.862
Table 2: Comparison of panorama and perspective teachers. Metrics are averaged equally over six cube faces for Stanford2D3D and four horizontal faces for Metropolis, as its top and bottom faces lack valid ground-truth depth. Best results are bolded.
Figure 2: Visualization of geometry-adaptive teacher weighting. Faces dominated by planar surfaces, such as walls or ceilings, receive higher panorama teacher weights, while views with greater geometric variation receive lower weights (blue or green).
Figure 3: Diagram of Spherical Cross-Level Feature Fusion. Cross-attention with a spherical direction bias is followed by a local feed-forward network with panorama padding and depthwise convolution. Reshaping is omitted for brevity.
Matterport3D
Stanford2D3D
Metropolis
Method
Params (M)
GFLOPs ↓
Latency (ms) ↓
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
Foundation models
DA360-S † [ 18 ]
24.804
299.520
517.412
0.168
0.802
0.082
0.947
0.232
0.597
DAP [ 23 ]
365.050
2543.697
1868.322
0.137
0.840
0.072
0.963
0.195
0.734
DepthMaster [ 46 ]
385.191
11521.707
OOM
0.101
0.905
0.062
0.965
1.022
0.263
PaGeR [ 5 ]
1286.563
26200.188
OOM
0.116
0.889
0.064
0.972
0.224
0.500
Table 3: Comparison with existing method without ground-truth scale alignment. † denotes models retrained or finetuned with our recipe. Latency is measured on a Jetson Orin NX 16GB in MAXN mode, with AMP FP16 and batch size 1. EG360 uses structural reparameterization. OOM denotes out of memory. Bold denotes the best value in each column within each group.
Figure 4: Qualitative comparisons on Matterport3D (top) and Metropolis (bottom). The first column shows ground-truth depth and RGB images; the remaining columns show predictions above per-pixel AbsRel maps for UniFuse, EG360, DA360-S, and DAP. Error colors range from purple (low) to yellow (high). Pixels without ground truth depth are masked in the error map.
Matterport3D
Stanford2D3D
Metropolis
Method
Params (M)
GFLOPs ↓
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
Baseline
6.144
21.196
0.199
0.722
0.126
0.850
0.266
0.578
+ APPD
6.144
21.196
0.169
0.781
0.101
0.903
0.244
0.619
+ SCFF
6.593
23.492
0.195
0.733
0.113
0.876
0.263
0.582
EG360
6.593
23.492
0.173
0.784
0.094
0.923
0.237
0.632
Table 4: Overall ablation of APPD and SCFF on three real-world test datasets. EG360 includes both components. Bold denotes the best value in each column.
Matterport3D
Stanford2D3D
Metropolis
Distill Method
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
No distill
0.199
0.722
0.126
0.850
0.266
0.578
DAP only
0.176
0.771
0.101
0.901
0.257
0.601
MoGe2 only
0.176
0.762
0.146
0.762
0.235
0.628
DAP + MoGe2 (Avg)
0.168
0.780
0.109
0.889
0.238
0.630
Table 5: Teacher complementarity with our baseline student. Using both teachers leads to better results.
Matterport3D
Stanford2D3D
Metropolis
Batch size
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
1
0.184
0.758
0.115
0.875
0.240
0.630
2
0.175
0.769
0.114
0.872
0.242
0.613
4
0.173
0.778
0.110
0.883
0.244
0.618
8
0.168
0.780
0.109
0.889
0.238
0.630
16
0.166
0.782
0.108
0.890
0.239
0.628
Table 6: Effect of the batch size for unlabeled data under Avg distillation. The first column is the unlabeled batch size.
Matterport3D
Stanford2D3D
Metropolis
Distill Method
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
Avg
0.168
0.780
0.109
0.889
0.238
0.630
Random
0.172
0.770
0.113
0.874
0.240
0.625
Manual
0.166
0.777
0.107
0.885
0.238
0.604
APPD
0.169
0.781
0.101
0.903
0.244
0.619
Table 7: Joint teacher strategies. Avg averages the two teacher losses. Random selects one teacher with equal probability per unlabeled batch. Manual assigns DAP to ZInD and MoGe2 to HoliCity. APPD is our geometry-adaptive loss weighting.
Matterport3D
Stanford2D3D
Metropolis
Loss
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
L1
0.168
0.780
0.109
0.889
0.238
0.630
SI
0.228
0.680
0.155
0.805
0.246
0.603
LSS
0.176
0.748
0.111
0.901
0.466
0.095
Table 8: Pseudo-label loss type ablation under Avg teacher fusion.
Matterport3D
Stanford2D3D
Metropolis
Method
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
AbsRel ↓
δ1↑
UniFuse
0.234
0.663
0.155
0.754
0.297
0.500
+ APPD
0.197
0.745
0.119
0.867
0.250
0.603
BiFuse++
0.295
0.555
0.204
0.717
0.398
0.260
+ APPD
0.211
0.718
0.158
0.760
0.270
0.546
ACDNet
0.233
0.670
0.146
0.795
0.301
0.487
Table 9: Applying APPD to other lightweight models.
Figure 5: Qualitative results of applying APPD to our baseline model. The error is reduced in this real-world example.
Figure 6: Qualitative results of using SCFF. The error for the large shelf is lowered as SCFF allows long-range context modeling.
Configuration
Params (M)
GFLOPs ↓
AbsRel ↓
δ1↑
Baseline
6.144
21.196
0.126
0.850
Full
6.593
23.492
0.113
0.876
Full − D
6.593
23.492
0.125
0.864
Full − Cross
6.296
21.814
0.119
0.867
Full −3×3 DW Conv
6.589
23.478
0.122
0.868
Full − Cross + C4 Self
6.839
22.173
0.119
0.875
Table 10: SCFF ablation on Stanford2D3D using labeled data only. Full is the complete SCFF module. − D removes the spherical direction bias, − Cross removes cross-level attention, and −3×3 DW Conv removes the depthwise convolution in the panoramic FFN. Full − Cross + C4 Self/C3 Self replaces the cross attention with a self-attention block on C4/C3.
Institute of Information Science, Beijing Jiaotong University, Beijing 100044, China. · Visual Intelligence +X International Cooperation Joint Laboratory of MOE, Beijing 100044, China.