TLT for custom dataset and deploy .engine file

Hi all,
I have some question about train the resnet18detectnetv2 for custom my own dataset.

1- I need to converted my own dataset to kitti dataset format? If so, I need first resize images and boxes for fixed size like kitti dataset size as offline?
2- for converting the dataset to tfrecord, tlt-tfrecord-converter expected to have 16 fields in labels text files, I should to fill other fields any value like zeros except x,y,w,h and class id?
3- In the detectnet_v2_train_resnet18_kitti.txt, there are all config about classes of kitti, like this:

  target_class_config {
    key: "pedestrian"
    value {
      clustering_config {
        coverage_threshold: 0.00749999983236
        dbscan_eps: 0.230000004172
        dbscan_min_samples: 0.0500000007451
        minimum_bounding_box_height: 20
      }
    }
  }

  target_classes {
    name: "pedestrian"
    class_weight: 4.0
    coverage_foreground_weight: 0.0500000007451
    objectives {
      name: "cov"
      initial_weight: 1.0
      weight_target: 1.0
    }
    objectives {
      name: "bbox"
      initial_weight: 10.0
      weight_target: 10.0
    }
  }

There are many such config about claases of kitti, I need to modify all of this to my custom dataset classes?

4- How do I use trt engine like .engine file with python code without deep stream sdk?

  1. You just need to convert your label to KITTI format. Reason: Is it possible to generate .tfrecorfs for tlt training directly without using intermediate kitti format?
    For detection network, you need to resize your images/labels to what resolution you want to train.
    See https://docs.nvidia.com/metropolis/TLT/tlt-getting-started-guide/index.html#requirements

  2. Not 16 fields. Should be 15 fields. Yes,except x1,y1,x2,y2 and class, just set other fields to 0 or 0.0. See 1) too.

  1. It has nothing to do with KITTI. Only you need to replace your class name.

  2. Reference:
    Deploy .engine file in python
    https://github.com/NVIDIA/object-detection-tensorrt-example/blob/master/SSD_Model/utils/inference.py