Collection Data for Detection is not training?

savunmasad · October 14, 2020, 11:37pm

Hi,

I collect my data with
camera-capture /dev/video1 and did same thing in this page:

dusty-nv/jetson-inference/blob/master/docs/pytorch-collect-detection.md

<img src="https://github.com/dusty-nv/jetson-inference/raw/master/docs/images/deep-vision-header.jpg" width="100%">
<p align="right"><sup><a href="pytorch-plants.md">Back</a> | <a href="../README.md#hello-ai-world">Next</a> | </sup><a href="../README.md#hello-ai-world"><sup>Contents</sup></a>
<br/>
<sup>Transfer Learning - Object Detection</sup></s></p>

# Collecting your own Detection Datasets

The previously used `camera-capture` tool can also label object detection datasets from live video:

<img src="https://github.com/dusty-nv/jetson-inference/raw/dev/docs/images/pytorch-collection-detect.jpg" >

When the `Dataset Type` drop-down is in Detection mode, the tool creates datasets in [Pascal VOC](http://host.robots.ox.ac.uk/pascal/VOC/) format (which is supported during training).

> **note:** if you want to label a set of images that you already have (as opposed to capturing them from camera), try using a tool like [`LabelImg`](https://github.com/tzutalin/labelImg) that also saves in Pascal VOC format.  If you need to label a video file, dump the video frames to images first.

## Creating the Label File

Under `jetson-inference/python/training/detection/ssd/data`, create an empty directory for storing your dataset and a text file that will define the class labels (usually called `labels.txt`).  The label file contains one class label per line, for example:

``` bash

This file has been truncated. show original

and training data with

python3 train_ssd.py --dataset-type=voc --data=data/kurucu --model-dir=models/kurucu --batch-size=2 --epochs=1

like this video that nvidia shared:

but it shows me this:

AastaLLL · October 15, 2020, 2:15am

Hi,

It seems the error message is cropped out in the screenshot.
Would you mind to share us a complete one?

Thanks.

savunmasad · October 15, 2020, 6:32am

no it’s not cropped out. After that terminal shows:
sadsavunma@savunma-desktop:/media/sadsavunma/SD/jetson-inference/python/training/detection/ssd$

this is what i’m writing here.

sadsavunma@sadsavunma-desktop:/media/sadsavunma/SD/jetson-inference/python/training/detection/ssd$ python3 train_ssd.py --dataset-type=voc --data=data/kurucu --model-dir=models/kurucu --batch-size=2 --workers=1 --epochs=1
2020-10-15 09:25:34 - Using CUDA...
2020-10-15 09:25:34 - Namespace(balance_data=False, base_net=None, base_net_lr=0.001, batch_size=2, checkpoint_folder='models/kurucu', dataset_type='voc', datasets=['data/kurucu'], debug_steps=10, extra_layers_lr=None, freeze_base_net=False, freeze_net=False, gamma=0.1, lr=0.01, mb2_width_mult=1.0, milestones='80,100', momentum=0.9, net='mb1-ssd', num_epochs=1, num_workers=1, pretrained_ssd='models/mobilenet-v1-ssd-mp-0_675.pth', resume=None, scheduler='cosine', t_max=100, use_cuda=True, validation_epochs=1, weight_decay=0.0005)
2020-10-15 09:25:34 - Prepare training datasets.
2020-10-15 09:25:34 - VOC Labels read from file: ('BACKGROUND', 'ataturk')
2020-10-15 09:25:34 - Stored labels into file models/kurucu/labels.txt.
2020-10-15 09:25:34 - Train dataset size: 24
2020-10-15 09:25:34 - Prepare Validation datasets.
Traceback (most recent call last):
  File "train_ssd.py", line 243, in <module>
    target_transform=target_transform, is_test=True)
  File "/media/sadsavunma/SD/jetson-inference/python/training/detection/ssd/vision/datasets/voc_dataset.py", line 24, in __init__
    self.ids = self._read_image_ids(image_sets_file)
  File "/media/sadsavunma/SD/jetson-inference/python/training/detection/ssd/vision/datasets/voc_dataset.py", line 91, in _read_image_ids
    with open(image_sets_file) as f:
sadsavunma@sadsavunma-desktop:/media/sadsavunma/SD/jetson-inference/python/training/detection/ssd$

dusty_nv · October 15, 2020, 4:27pm

Can you check if the ImageSets/Main directory of your dataset contains the following files?

test.txt train.txt trainval.txt val.txt

If you are missing some of those, can you replicate the file that does exist? For example, cp train.txt test.txt, ect.

savunmasad · October 15, 2020, 4:45pm

Does not contain test.txt but others dont remember and i tried today for classification the datasets but same problem happened but now I can’t give the error because I don’t use it.

dusty_nv · November 26, 2020, 12:44pm

OK, my guess is that you may have train.txt, but not the others - so copy train.txt so you also have test.txt trainval.txt val.txt. I will have to add a better error message about this in the future.

The classification mode of the camera-capture tool saves the datasets in a structure for classification training only. And the detection mode of camera-capture tool saves the datasets in a structure for detection training.

savunmasad · October 15, 2020, 5:02pm

Ok, thank you. I will try tomorrow

savunmasad · November 26, 2020, 12:49pm

now, when run the camera-capture /dev/video1
terminal shows:
camera-capture: failed to capture RGBA image from camera
but Data Capture Control window is coming

My camera is ZED, and i don’t know how to use Raspberry Pi camera for jetson tx2, how to plug to where of Jetson TX2

dusty_nv · November 30, 2020, 7:19pm

Hmm I haven’t tested this with the ZED, but I do know that the ZED has both it’s left/right cameras in the same video stream, so I’m not sure that would work well in this circumstance for data annotation. The ZED would require extra pre-processing to crop out either the right or left frame.

Are you using a Jetson TX2 Developer Kit? If so, the TX2 devkit comes with an onboard MIPI CSI camera already (based on OV5693 sensor). You can try to view it by running video-viewer csi://0

savunmasad · November 30, 2020, 7:30pm

Yes TX2 developer kit and unfortunately onboard camera is not moving :)

dusty_nv · November 30, 2020, 7:48pm

If you run nvgstcapture-1.0 program, do you get a camera feed from the onboard camera? Does it show any image?

If your onboard camera isn’t working, you might want to try a monocular USB webcam like the Logitech C270 or C920 (or similar off-brand USB webcams can work too).

savunmasad · November 30, 2020, 8:17pm

Onboard camera, usbwebcam and zed working clearly.
The second camera of the zed camera freezes when it is running, and when I do v4l2, the entire resolution for zed is not listed only.
there is no other problem.