karasawatakumi / monodepth-dev

Monocular depth estimation by segmentation models for NYU-depth v2 dataset. Simple PyTorch Lightning implementation.

Monocular Depth Estimation for NYU2

This repository provides a simple PyTorch Lightning implementation of monocular depth estimation for NYU Depth Dataset V2.

Dependencies

• Docker 20.10.2
• Docker Compose 1.28.3
• Python 3.8.0
• Segmentation Models 0.1.3
• PyTorch 1.6.0
• PyTorch Lightning 1.2.5
• OmegaConf 2.0.6
• Weights & Biases 0.10.25

Please see requirements.txt for the other libraries' versions.

Approach

Segmentation model + *Depth loss

*We use the three loss proposed in [J. Hu+, 2019].

BACKBONE	TYPE	delta1	delta2	delta3	lg10	abs_rel	mae	mse
efficientnet-b7	UnetPlusPlus	0.8381	0.9658	0.9914	0.0553	0.1295	0.3464	0.3307
efficientnet-b7	FPN	0.8378	0.9662	0.9915	0.0561	0.1308	0.3523	0.3308
efficientnet-b4	UnetPlusPlus	0.8361	0.9649	0.9913	0.0559	0.1308	0.3488	0.3293
efficientnet-b4	Unet	0.8312	0.9636	0.9905	0.0569	0.1321	0.3582	0.3508
efficientnet-b4	FPN	0.8308	0.9648	0.9909	0.0570	0.1337	0.3581	0.3411
efficientnet-b4	DeepLabV3Plus	0.8304	0.9634	0.9900	0.0570	0.1352	0.3596	0.3483
resnet50	FPN	0.8287	0.9637	0.9905	0.0577	0.1351	0.3600	0.3456
resnet50	Unet	0.8277	0.9619	0.9903	0.0576	0.1345	0.3570	0.3421
resnet50	MyUnet	0.8273	0.9612	0.9894	0.0577	0.1343	0.3576	0.3458
resnet50	UnetPlusPlus	0.8241	0.9623	0.9896	0.0581	0.1356	0.3610	0.3486
resnet50	DeepLabV3Plus	0.8225	0.9608	0.9888	0.0583	0.1375	0.3639	0.3569
efficientnet-b0	UnetPlusPlus	0.8190	0.9607	0.9894	0.0592	0.1396	0.3722	0.3667
efficientnet-b0	FPN	0.8132	0.9597	0.9897	0.0601	0.1415	0.3780	0.3728

NOTE: To simplify the experiment, we set the image size to [288, 224] (divisible by 32), which is not exactly the same as the evaluation in the paper.

Preparation

Dataset: NYU Depth Dataset V2

sh scripts/prepare_nyu2.sh

This script uses the downloading link in J. Hu's repository.

Installation

docker-compose build
docker-compose run dev

Run

Train

python tools/train.py

usage: train.py [-h] [--config CONFIG] [--resume RESUME] [--gpu-ids GPU_IDS [GPU_IDS ...] | --n-gpu N_GPU] [--amp {O1,O2,O3}]
                [--profiler {simple,advanced}]
                ...

Train a predictor

positional arguments:
  opts                  Overwrite configs. (ex. OUTPUT_DIR=results, SOLVER.NUM_WORKERS=8)

optional arguments:
  -h, --help            show this help message and exit
  --config CONFIG       Optional config path. `configs/default.yaml` is loaded by default
  --resume RESUME       the checkpoint file to resume from
  --gpu-ids GPU_IDS [GPU_IDS ...]
  --n-gpu N_GPU
  --amp {O1,O2,O3}      amp opt level
  --profiler {simple,advanced}
                        'simple' or 'advanced'

If you want to override the config with command line args, put them at the end in the form of dotlist.

python tools/train.py --config [config path] SOLVER.NUM_WORKERS=8 SOLVER.EPOCH=5

Credit

@inproceedings{Hu2019RevisitingSI,
  title={Revisiting Single Image Depth Estimation: Toward Higher Resolution Maps With Accurate Object Boundaries},
  author={Junjie Hu and Mete Ozay and Yan Zhang and Takayuki Okatani},
  journal={2019 IEEE Winter Conference on Applications of Computer Vision (WACV)},
  year={2019}
}