Clarification about ReIdentificatioNet pre-training

[corentin87]

Hi,

For ReIdentificationNet, a pre-trained model is available in the NGC catalog trainable_v1.1: resnet50_market1501_aicity156.tlt . I can’t find many information about the pre-trained process though.

In this great blogpost https://developer.nvidia.com/blog/enhance-multi-camera-tracking-accuracy-by-fine-tuning-ai-models-with-synthetic-data/, they are saying the network has been pre-trained using SOLIDER technic and on a dataset which:

includes a combination of NVIDIA proprietary datasets along with Open Images V5.

However in the NGC page, they are only mentioning market-1501 + synthetic IDs. Therefore, I am wondering on which dataset has the model been pre-trained? is it on a combination of NVIDIA proprietary datasets along with Open Images or is it on market-1501 + synthetic data?

Also the blog mention, fine-tuning using 4470 real IDs, does that mean the model tested here is different from deployable_v1.2 for NGC?

Thank you

[Morganh]

@corentin87

does that mean the model tested here is different from deployable_v1.2 for NGC?

The deployable_v1.2 is trained with 14737 images from Market-1501 dataset of 751 real people and 29533 images of 156 people (148 of which are synthetic) from MTMC people tracking dataset from the 2023 AI City Challenge.

@corentin87

Therefore, I am wondering on which dataset has the model been pre-trained?

The training dataset is mentioned in ngc model page. Refer to ReIdentificationNet | NVIDIA NGC.

[corentin87]

Thanks for the answer. However it is still not clear to me. Your two replies are mentioning the same information

Therefore:

• is deployable_v1.2 fine-tune from trainable_v1.1? or is it the export version?
• If they are the same, what is the pre-trained model used to train trainable_v1.1?
• what about the the pre-trained model describe in the blog?

[Morganh]

Yes, it is the export version. The models in model card are trained on Market1501 + synthetic datasets.

The pretrained-model mentioned in the blog or mentioned in notebook tao_tutorials/notebooks/tao_launcher_starter_kit/re_identification_net/reidentificationnet_swin.ipynb at main · NVIDIA/tao_tutorials · GitHub is trained on unlabeled datasets. We train it on ~3M image unlabeled crops of people objects. Reidentificationnet_transformer is the network trained on combination of NVIDIA proprietary datasets along with Open Images V5 (~3M images).

[corentin87]

Ok thanks for the reply.

I have another question regarding the training output displayed
For the training, I have :

Subset   │   # IDs │   # Images │   
╞══════════╪═════════╪
│ Train    │     162 │      41712

and I have configured:

num_classes: 162
batch_size: 128
val_batch_size: 64
num_workers: 4
num_instances: 4

Therefore, we should have 41712/128 = 236 batches for each epoch. I am wondering why TAO is showing 570?
And why are the epochs split into 2? thought this was due to num_workers but even setting it up to 1 shows a split epochs output.

Thanks

[Morganh]

@corentin87

showing 570

Should be related to validation dataset. It contains two parts. One is training part, another is validation part.