Event cameras keep capturing structure where conventional cameras blur or saturate, which makes them attractive for 3D perception in fast or poorly lit scenes. Event datasets with dense 3D ground truth, however, are far smaller and less diverse than the image collections behind feed-forward 3D reconstruction models. We present Déjà-Events, an event-only model that predicts depth, camera pose and pointmaps from multiple views in a single forward pass. Rather than relying on scarce 3D labels, we distil it from Déjà View, a feed-forward RGB model that refines its tokens by repeatedly applying one shared transformer block. On paired RGB and event recordings, the event student learns to reproduce the frozen teacher's final tokens, which requires no depth or pose ground truth and trains only about 1% of its parameters. Because events are dense at edges and moving texture but nearly absent in textureless regions, we further introduce sparsity-aware adaptive gating, which adapts the shared block's gates to events and adjusts them per patch by event density, so that accuracy holds as refinement steps are added. Trained only on simulated data, Déjà-Events achieves the lowest pointmap error among the event-based methods we evaluate, on EventScape and zero-shot on the real MVSEC and DSEC datasets. On DSEC it halves the error of the strongest baseline (0.15 vs. 0.30 AbsRel) with about a tenth of its parameters. It also has the lowest depth error on EventScape and DSEC and the lowest trajectory error on EventScape.
Figures & tables
Fig. 1: Qualitative results on PEOD [ 10 ] . At a tunnel exit the RGB frame is saturated across most of the road and the vehicles ahead, while the event stream retains structure. In the blown-out area the RGB teacher predicts a flat, ground-plane-like depth, while the event student keeps the vehicles and the tunnel structure. Neither model was trained on PEOD.
Fig. 2: Overview. A frozen RGB DVLT teacher (top) and an event student with the same architecture (bottom) process aligned RGB frames and events. In each training iteration both run the shared block for the same sampled number of steps K , and the loss matches their final tokens (Sec. III-B ). The student trains low-rank adapters in DINOv2 and in the shared block, and its gates are unfrozen and corrected per patch in both attention blocks (Sec. III-C ).
EventScape
MVSEC
DSEC
Model
AbsRel ↓
Δ1↑
AbsRel ↓
Δ1↑
AbsRel ↓
Δ1↑
DVLT teacher
0.111
0.898
0.290
0.556
0.060
0.965
Déjà-Events
0.144
0.843
0.298
0.608
0.145
0.841
EventVGGT
0.162
0.751
0.275
0.618
0.282
0.550
E2Depth
0.283
0.543
0.381
0.417
0.408
0.416
EReFormer
0.241
0.648
0.411
0.421
0.409 ∗
0.511 ∗
TABLE I: Depth estimation.
Fig. 4: Zero-shot depth on DSEC. Frame 685 of interlaken_00_b , from a 50 ms event window. All columns show the identical 476×630 crop of DSEC’s event camera. This is a DSEC test-split sequence, which ships no public disparity, so the comparison is qualitative; no model was trained on DSEC and all baseline checkpoints are EventScape-trained. Each depth map is colourised against its own 1st/99th percentile.
EventScape
MVSEC
DSEC
Model
ATE ↓
AUC 20 ↑
ATE ↓
AUC 20 ↑
ATE ↓
AUC 20 ↑
DVLT teacher
0.013
0.953
0.144
0.732
0.043
0.897
Déjà-Events
0.036
0.887
0.095
0.717
0.105
0.811
EventVGGT
0.796
0.029
1.327
0.031
3.785
0.018
DEVO
0.062
0.850
0.034
0.731
0.027
0.972
TABLE II: Camera pose.
EventScape
MVSEC
DSEC
Model
AbsRel ↓
Inl@25 ↑
AbsRel ↓
Inl@25 ↑
AbsRel ↓
Inl@25 ↑
DVLT teacher
0.187
0.794
0.350
0.452
0.058
0.976
Déjà-Events
0.202
0.789
0.297
0.575
0.148
0.855
EventVGGT A
0.356
0.451
0.347
0.514
0.301
0.577
EventVGGT B
0.700
0.201
0.459
0.368
0.331
0.538
modular (DEVO pose + GT intrinsics)
TABLE III: Pointmap reconstruction.
Fig. 5: Pointmap from the 18-view window of interlaken_00_b containing the frame of Fig. 4 , predicted by our student from events alone; points are coloured with the aligned RGB frames for legibility, and the frustums are the student’s predicted cameras.
Depth
Pose
Pointmap
Model
AbsRel ↓
Δ1↑
ATE ↓
AUC 20 ↑
AbsRel ↓
DVLT teacher
0.339
0.507
0.156
0.823
0.424
Déjà-Events
0.373
0.565
0.065
0.790
0.347
Déjà-Events + day1
0.324
0.567
0.066
0.821
0.418
TABLE IV: MVSEC outdoor_night1.
GFLOPs
Peak mem.
Model
Params
Trained
/ frame ↓
(GiB) ↓
DVLT teacher
117.1
–
597.7
1.62
Déjà-Events
118.4
1.60
604.6
1.65
EventVGGT A
1161.1 ‡
3.15 †
2209.0
4.21 ‡
EventVGGT B
1159.5
1.57 †
1265.0
4.20
DEVO + E2Depth
14.1
14.1
628.7
0.79
TABLE V: Model size and compute.
Student configuration
AbsRel ↓
Δ1↑
ATE ↓
DVLT teacher (RGB, ceiling)
0.111
0.898
0.013
Distillation only, frozen schedule
0.147
0.839
0.036
+ unfrozen schedule
0.146
0.841
0.036
+ SAAG (ours)
0.144
0.843
0.036
TABLE VI: Step-scale ablation.
Fig. 6: (a) Per-step correlation between the token update Δzk and event density, for the frozen-schedule student, the same student with an unfrozen schedule, our final patchwise correction, and an RGB teacher probed on the same clips as a control. (b) AbsRel on DSEC (zero-shot) as inference steps K are swept past the [8,16] range seen in training.