Weak performance SDD ,detect person TLT training


I try to do the training for the detection of people I use the SSD as a model, on the other hand I always have a weak mAP around 0.0005 ~ 0.000009
I changed my database but it’s still the same performance, I thought maybe the network “resnet_18.hdf5”
I’m working with this tutorial: SSD — Transfer Learning Toolkit 3.0 documentation

any suggestion please ,

I suggest you running the SSD jupyter notebook firstly. It trained the public KITTI dataset.
Then you can replace it with your own dataset.

in fact, I use the data “The KITTI Vision Benchmark Suite” the kitti format already present?
I don’t know where this low mAP value comes from ?

Did you use the correct pretrained model for SSD? See Overview — Transfer Learning Toolkit 3.0 documentation

More, as mentioned above, please download the jupyter notebook and run with it.

all the configuration has been done, in the same way I did the training for the detection of vehicles it worked well with an accuracy of 86%, but when I changed the class to the person the precision is very low

In default SSD jupyer notebook, if you train with KITTI dataset with default training spec, it will get the AP of person class. Right?

yes mAP pedestrian = 0.0005

Can you attach the .ipynb file here?

os.system("/home/sylia/.local/bin/tlt ssd train -e /workspace/tlt -experiments/Data/Work/resnet18/config/TrainEvaluator.txt -r /workspace/tlt-experiments/Data/Work/resnet18/ResultTrain/ -k KEY -m /workspace/tlt-experiments/Data/Work/resnet18/pretrained_resnet18/tlt_resnet18_ssd_v1/resnet_18.hdf5 --gpus 1")

this is config file
TrainEvaluator.txt (1.4 KB)

I work on the terminal

As we discussed above, you already training with SSD jupyter notebook against KITTI dataset with default training spec file, right?
Please share the .ipynb file here.

I do not use a jupyter notebook file because I run the command line in the terminal, indeed, I have respected all the requirements described in the jupyter notebook

If possible, please share the full log.

  • Prepare dataset and pre-trained model
  • Run training

for data :

The KITTI Vision Benchmark Suite. the KITTI detection images (Download) and labels (Download)

for model :

ngc registry model download-version nvidia/tlt_pretrained_object_detection:resnet18

Run traininig :

tlt ssd train -e /workspace/tlt-experiments/Data/Work/resnet18/config/TrainEvaluator.txt -r /workspace/tlt-experiments/Data/Work/resnet18/ResultTrain/ -k KEY -m /workspace/tlt-experiments/Data/Work/resnet18/pretrained_resnet18/tlt_resnet18_ssd_v1/resnet_18.hdf5 --gpus 1

config file for training:
ssd_train_resnet18_kitti.txt (1.4 KB)

Is there any training log?

training log is like that of cars just at the level of the mAP have very low values in the training for people ~ 0.0008 0.0005

I will run training with SSD jupyter notebook against KITTI dataset with default training spec file. To check if I can reproduce your error.
Please make sure we are on the same page.


I train SSD network with KITTI dataset. Modify epoch to 20. The AP for person is not 0.0005 ~ 0.000009.
I attach training log and spec. Please retry on your side.
SSD_training_KITTI_300_300.txt (57.5 KB)

when I do the training just for one class (person) I get this

but when it’s the 3 classes I have this

There is no update from you for a period, assuming this is not an issue any more.
Hence we are closing this topic. If need further support, please open a new one.

Please share your full log for running 3 classes.