Not able to achieve the performance metrics from paper for offline training with phi = 3

[yashj235]

Hi,

I am running the code with offline_train_full_config_phi3_BN.yaml. Only change I make is train it on a single GPU instead of multiple. However, the metrics that I achieve on lightbox and sunlamp are different from the reported ones.

Run	lightbox eR (heatmap) [deg]	lightbox eT (efficientpose) [m]	lightbox Final Pose	sunlamp eR (heatmap) [deg]	sunlamp eT (efficientpose) [m]	sunlamp Final Pose
run1	8.11799	0.19787	0.17407	14.93889	0.25372	0.30487
run2	7.74038	0.20117	0.16770	15.16580	0.27020	0.31122
run3	8.83201	0.25162	0.19437	14.98225	0.29293	0.30930
average	8.23013	0.21689	0.17871	15.02898	0.27228	0.30846

For generating style augmentation, my pipeline is as follows:

1. Calculate the mean and variance for speed+ synthetic images using get_embedding_mean_and_covariance.py
2. Create style augmented images for the synthetic images
3. Convert them to greyscale and save

Am I missing something? Kindly suggest on how to improve the performance further.