GuidedDecoding

Accompaning repository for the 2022 ICRA paper "Lightweight Monocular Depth Estimation through Guided Decoding"

Trained weights

Dataset	Resolution	Model-Version
NYU Depth V2	240x320 (Half)	GuideDepth
NYU Depth V2	240x320 (Half)	GuideDepth-S
NYU Depth V2	480x640 (Full)	GuideDepth
NYU Depth V2	480x640 (Full)	GuideDepth-S

Evaluation procedure (on GPU)

For the evaluation, download the already prepared testsets from here:

NYU Depth V2

Unpack the data for Evaluation

run main.py --eval --dataset DATASET --resolution RESOLUTION --model MODEL_NAME --test_path PATH_TO_TEST_DATA --num_workers=NUM_WORKERS --save_results PATH_TO_RESULTS

You can select from the following options:

[RESOLUTION: full, half]
[DATASET: nyu_reduced]

Inference and deployment

We performed our evaluation on the NVIDIA Jetson Nano and the NVIDIA Xavier NX, using the following dependencies:

Jetpack: 4.5.1

CUDA: 10.2

CUDNN: 8.0.0

Python: 3.6.9

tensorRT: 7.1.3

PyTorch: 1.8.0

torchvision: 0.9.1

torch2trt: 0.2.0

Installing PyTorch and torchvision, refer to this post: https://forums.developer.nvidia.com/t/pytorch-for-jetson-version-1-11-now-available/72048

Installing torch2trt: https://github.com/NVIDIA-AI-IOT/torch2trt

You might need to increase SWAP memory for the tensorRT conversion to 4GB: https://github.com/JetsonHacksNano/resizeSwapMemory

Usage

python3 inference.py --eval --model MODEL_NAME --resolution RESOLUTION --dataset DATASET --weights_path PATH_TO_WEIGHTS --save_results PATH_TO_RESULTS --test_path PATH_TO_TEST_DATA

By selecting from the following options:

[RESOLUTION: full, half]
[DATASET: nyu_reduced, ]

Training

You will need the pretrained weights for DDRNet-23 slim, which can be downloaded here or acquired from the official repository

Preparing NYU Depth V2

We used a Subset of NYU Depth V2 designed and prepared by Alhashim et al. (https://github.com/ialhashim/DenseDepth)

To train, download the dataset linked in their repository. No need to unpack, the dataloader loads the compressed data.

Training procedure

run main.py --train --dataset DATASET --resolution RESOLUTION --model MODEL_NAME --data_path PATH_TO_TRAINING_DATA --num_workers=NUM_WORKERS --save_checkpoint PATH_TO_CHECKPOINTS

duraes-antonio/seg_depth