yaoyao-liu/SyntheticNeMo

Synthetic image generation from 3D model and NeMo pose estimation.

SyntheticNeMo

Synthetic image generation from 3D and NeMo pose estimation.

Generate Synthetic Data from 3D

See: Instruction for Generating Synthetic Data

NeMo Pose Estimation Training on Synthetic

See training script: pretrain.sh

And run

cd train
. pretrain.sh

NeMo Pose Estimation Testing on Real

See test script: test.sh

And run

cd test_pascal3d
. test.sh

NeMo Pose Estimation Training: Synthetic + Pseudo Labeling on Real

After finishing these steps:

1. NeMo Pose Estimation Training on Synthetic

2. NeMo Pose Estimation Testing on Real with additional argument "--prepare_for_pseudo" for each command line.

See script for pseudo labeling: pseudo.sh

And run

cd train
. pseudo.sh

NeMo Pose Estimation Training: Synthetic + Fine-tune on Real

After finishing these steps:

1. NeMo Pose Estimation Training on Synthetic

2. NeMo Pose Estimation Testing on Real with additional argument "--prepare_for_pseudo" for each command line.

See script for fine-tuning: fine_tune.sh

And run

cd train
. fine_tune.sh

You can adjust the number of images used for fine-tuning by modifying --num_samples.

By default, --num_samples 92 means fine-tuning on 10% of the real data.