01⌁
Black-box
prediction
Use logits exposed by the frozen source predictor. Adaptation never accesses source examples or internal parameters.
SOURCE-FREE
WACV 2027Core A
3D-BBDA: Backpropagation-Free Black-BoxDomain Adaptation for Robust 3D Point Cloud Recognition
Department of Data Science and Engineering,
Indian Institute of Science Education and Research Bhopal, India
{shivangir23, kunal24, vaibhav, vinodkk}@iiserb.ac.in
*Equal contribution.
FIGURE 1 · METHOD OVERVIEWBLACK-BOX / SOURCE-FREE ADAPTATION
Figure 1: 3D-BBDA routes unlabeled target data through a non-parametric network and combines Gaussian-perturbed features, feature prototypes, subspace alignment, and a weighted prediction module—without updating the source-trained model.
01 ABSTRACT
Robustness without
access.
Robust 3D point-cloud recognition is critical for robotics, mapping, and urban sensing, yet classifiers often suffer performance drops under sensor changes, environmental variation, and acquisition artifacts.
We present 3D-BBDA, a backpropagation-free black-box domain adaptation framework that adapts a source-trained predictor to unlabeled target data using only black-box outputs. Our method performs pseudo-label-guided prototype-based alignment and employs a weighted prediction module to fuse source logits with alignment logits, reducing sensitivity to noisy pseudo-labels.
No source data. No model parameters. No gradient updates.
The deployment constraint. Source data and the predictor are locked; only unlabeled target data is available at adaptation time.
02 HOW IT WORKS
Make the model
the only thing you need.
Three ideas turn a restricted deployment setting into a practical adaptation loop.
Black-box
02◎
Pseudo-label
alignment
Build class prototypes from target predictions and align target representations to stabilize the domain shift.
TARGET ONLY
03✦
Robust
prototype memory
Gaussian perturbations create a fallback memory for missing or unreliable class prototypes during alignment.
NO GRADIENTS
03 EXPERIMENTAL SIGNAL
Built for shifts
that break assumptions.
Evaluated on corrupted 3D benchmarks to test whether a model can stay useful when its original world changes.
32.08%mean error · PointNet
29.62%mean error · GDANet
35.88%source-only · PointNet
34.03%source-only · GDANet
ModelNet-40C · 4 corruption benchmarksLower error is better ↓
04 RESOURCES & CITATION
Read the work.
Reproduce the idea.
BibTeX
@inproceedings{3dbbda2027,
title = {3D-BBDA: Backpropagation-Free Black-Box Domain Adaptation for Robust 3D Point Cloud Recognition},
author = {Rai, Shivangi and Jangid, Kunal and Kumar, Vaibhav and Kurmi, Vinod K.},
booktitle = {WACV}, year = {2027}
}