README.md

BiCross - Unsupervised Spike Depth Estimation via Cross-modality Cross-domain Knowledge Transfer

Architecture

Result

Synthetic to Real

Extreme Weathers

Scene Changes

Respike

Usage

Train the model via BiCross

python train.py

• Training stages:
  1. Since the pre-trained parameters of DPT are trained on the ImageNet when you train from scratch, please first pretrain the model on the source RGB to adapt to the depth estimation task, changing the stage option in the train_config.json to pretrain and training for about 30 epochs.
  2. After the pre-train stage, set stage in train_config.json to crossmodality and continue training for another 10 epochs from source RGB to source spike.
  3. Finally, set stage in train_config.json to crossdomain and then continue training for about 20 epochs from source spike to target spike.

Test the trained model

python test.py

Visualize the results

python visualize.py

You can modify the configs for different training and testing configurations.

Visualization

BiCross Feature

The Coarse-to-Fine Knowledge Distillation (CFKD) module is proposed to transfer open-source RGB depth labels and sufficient semantic knowledge to mediate the simulated source spike domain. The global image-level branch and local pixel-level branch aim to replenish the sparse spatial-wise feature representation in spike and offer more feature-level information to align in cross-domain. With the help of CFKD, the semantic representation ability of spike features in different positions of the decoder is improved, showing similar results with RGB features.

The Self-Correcting Teacher-Student mechanism (SCTS) module is proposed to exploit coarse branch (in CFKD) transferred global feature to coherent align in the cross-domain stage. Meanwhile, the Uncertainty Guided Teacher-Student (UGTS) module adopts the mean-teacher mechanism to fully exploit the unlabeled target domain spike data, and the uncertainty-guided mechanism which abandons unreliable depth prediction to avoid error accumulation. With the help of both GLFA and UGMT, the perception and representation abilities of target spike data are improved.

Qualitative Result

Here, we present the qualitative results of BiCross in three scenarios: Synthetic to Real, Extreme Weathers, and Scene Changes.

Demo