SSD MobileNet V2 recipe

[Budelon]

💡 Your Question

I've read the docs on the COCO SSD recipe. The recipe train the model with

• DetectionHSV: prob: 1.0 # probability to apply HSV transform hgain: 5 # HSV transform hue gain (randomly sampled from [-hgain, hgain]) sgain: 30 # HSV transform saturation gain (randomly sampled from [-sgain, sgain]) vgain: 30 # HSV transform value gain (randomly sampled from [-vgain, vgain])

But it validates without the HSV transform. val_dataset_params: data_dir: /data/coco # root path to coco data subdir: images/val2017 # sub directory path of data_dir containing the train data. json_file: instances_val2017.json # path to coco train json file, data_dir/annotations/train_json_file. input_dim: [320, 320] cache_annotations: True ignore_empty_annotations: True transforms:

• DetectionPaddedRescale: input_dim: ${dataset_params.val_dataset_params.input_dim}
• DetectionTargetsFormatTransform: input_dim: ${dataset_params.val_dataset_params.input_dim} output_format: LABEL_NORMALIZED_CXCYWH class_inclusion_list: max_num_samples: with_crowd: True

Questions:

1. Why does the validation not use the HSV transform?
2. If i want to run the model trained on COCO, should i transform the BGR image to HSV before running?

Thanks.

Versions

No response

[BloodAxe]

There is no mistake.A DetectionHSV is a augmentation transformation, which takes image in BGR colorspace, transform it to HSV colorspace and apply random shifts to H,S and V components and then convert image back to BGR. This is data augmentation transformation aimed to improve model robustness to color changes. On validation you don't want to have any sort of randomness.