Depth Prediction¶
MXNet Pytorch
MXNet¶
Here is the model zoo for the task of depth prediction.
Hint
Training commands work with this script:
Download train.py
The test script Download test.py
can be used for
evaluating the models on various datasets.
KITTI Dataset¶
The following table lists pre-trained models trained on KITTI.
Hint
The test script Download test.py
can be used for
evaluating the models (KITTI RAW results are evaluated using the official server). For example
monodepth2_resnet18_kitti_stereo_640x192
:
python test.py --model_zoo monodepth2_resnet18_kitti_stereo_640x192 --pretrained_type gluoncv --batch_size 1 --eval_stereo --png
Note
Our pre-trained models reproduce results from recent state-of-the-art approaches. Please check the reference paper for further information.
Modality
is the method used during training. Stereo
means we use stereo image pairs to calculate the loss, Mono
means we use monocular image sequences to calculate the loss,
Mono + Stereo
means both the stereo image pairs and monocular image sequences are used to calculate the loss.
Resolution
is the input size of the model during training. 640x192
means we resize the raw image (1242x375) to 640x192.
Name |
Modality |
Resolution |
Abs. Rel. Error |
delta < 1.25 |
Hashtag |
Train Command |
Train Log |
---|---|---|---|---|---|---|---|
monodepth2_resnet18_kitti_stereo_640x192 1 |
Stereo |
640x192 |
0.114 |
0.860 |
83eea4a9 |
||
monodepth2_resnet18_kitti_mono_640x192 1 |
Mono |
640x192 |
0.121 |
0.858 |
c881771d |
||
monodepth2_resnet18_kitti_mono_stereo_640x192 1 |
Mono + Stereo |
640x192 |
0.109 |
0.872 |
9515c219 |
PoseNet¶
Monodepth2 trains depth and pose models at the same time via a self-supervised manner. So, we also give reproduced results of our pre-trained models here.
Hint
The test script Download test_pose.py
can be used for
evaluating the models (KITTI Odometry results are evaluated using the official server). For example
monodepth2_resnet18_posenet_kitti_mono_stereo_640x192
:
python test_pose.py --model_zoo_pose monodepth2_resnet18_posenet_kitti_mono_640x192 --data_path ~/.mxnet/datasets/kitti/kitti_odom --eval_split odom_9 --pretrained_type gluoncv --batch_size 1 --png
Please check the full tutorials Testing PoseNet from image sequences with pre-trained Monodepth2 Pose models.
Note
Our pre-trained models reproduce results from recent state-of-the-art approaches. Please check the reference paper for further information.
Sequence 09
and Sequence 10
means the model is tested on sequence 9 and sequence 10 of the KITTI Odometry dataset respectively.
Results show the average absolute trajectory error (ATE), and standard deviation, in meter.
Name |
Modality |
Resolution |
Sequence 09 |
Sequence 10 |
---|---|---|---|---|
monodepth2_resnet18_posenet_kitti_mono_640x192 1 |
Mono |
640x192 |
0.021±0.012 |
0.018±0.011 |
monodepth2_resnet18_posenet_kitti_mono_stereo_640x192 1 |
Mono + Stereo |
640x192 |
0.021±0.010 |
0.016±0.010 |
PyTorch¶
Models implemented using PyTorch will be added later. Please checkout our MXNet implementation instead.