bosch-ros-pkg/bstld

Sample scripts for the Bosch Small Traffic Lights Dataset

Bosch Small Traffic Lights Dataset

This repository contains some scripts to get started with the Bosch Small Traffic Lights Dataset (BSTLD). Contributions are very welcome. Simply create a pull request.

Dataset

The dataset can be downloaded here. A preview of the dataset is available on YouTube by clicking on the image.

Instructions on how to unzip *.zip.00X files can, for example, be found at https://hiro.bsd.uchicago.edu/node/3168 Update label files are in the label_files folder.

To convert Bosch Small Traffic Lights Dataset Annotations to Pascal VOC Format

 python bosch_to_pascal.py input_yaml out_folder

Sample Detections

A sample detection based on an adapted Yolo v1 model run on crops can be viewed at ]

Results

Method	Execution time	weighted mAP	mAP	Off	Green	Yellow	Red	External data	Link
Baseline	<100 ms	0.36						no	https://ieeexplore.ieee.org/document/7989163/
Hierarchical Deep Architecture	~150 ms	0.53						no	https://arxiv.org/abs/1806.07987
SSD Mobilenet V1	38 ms	0.60	0.41	0.00	0.68	0.41	0.55	no	https://github.com/bosch-ros-pkg/bstld/blob/master/tf_object_detection/configs/ssd_mobilenet_v1.config
Faster RCNN NAS-A	~1560s	0.65	0.43	0.00	0.71	0.33	0.66	no	https://github.com/bosch-ros-pkg/bstld/blob/master/tf_object_detection/configs/faster_rcnn_nas.config

Values are self-reported. The evaluation is performed on the test-set without empty frames. For different goals, e.g. using minimal training data, using external data only, or others, new tables can be created. We specifically encourage non-conventional approaches. Please make sure not to incorporate the test-set into your training, which includes multiple evaluations for different checkpoints of the same method. We understand that there can be larger variations between the different class average precisions, specifically due to the biased distribution. We will try to incorporate variations in results of the same method once reported.

Citation

In case of publication based on this dataset, please cite
@inproceedings{behrendt2017deep,
  title={A deep learning approach to traffic lights: Detection, tracking, and classification},
  author={Behrendt, Karsten and Novak, Libor and Botros, Rami},
  booktitle={Robotics and Automation (ICRA), 2017 IEEE International Conference on},
  pages={1370--1377},
  year={2017},
  organization={IEEE}
}

bstld's Issues

No Instruction on Inference.py

Hello

I am a student trying to recreate a research paper for which I need to use bstld datset trained SSD architecture with Mobilenet_v1. However inference.py lacks instructions on how to use it to get a pre-trained model.

Dataset download fails for dataset_train_rgb.zip.001-003 on chrome or firefox

Hello,

I am trying to download the training dataset but I am encountering an error while doing so.

I've registered my email in the form and follow the given link to download these files:

• dataset_train_rgb.zip.001
• dataset_train_rgb.zip.002
• dataset_train_rgb.zip.003
• dataset_train_rgb.zip.004

I couldn't download the 1, 2 and 3 as I encountered these problems:

• Chrome (Windows & Linux) Once I reached 1.7 Gb / 1.9 Gb the download restarted at 0 again and again
• Firefox (Windows & Linux) Once I reached 1.7 Gb / 1.9 Gb the download is marked as failed

However, I succeed to download them on Windows with Internet Explorer.

I hope it's the right place to report this issue as I didn't find emails on the dataset page.

Best regards

'IndexError: list index out of range' when running the to_tfrecords.py file

Hi, all
Following the 'Tensorflow object detection example' RAEDME, I got an error like this when compiling the to_tfrecords.py file

Traceback (most recent call last):
File "D:\Udacity\Self-driving_car\projects\Term3\P3\models\research\bstld\tf_object_detection\to_tfrecords.py", line 195, in
create_datasets(config)
File "D:\Udacity\Self-driving_car\projects\Term3\P3\models\research\bstld\tf_object_detection\to_tfrecords.py", line 149, in create_datasets
train_labels, valid_labels = split_train_labels(train_labels)
File "D:\Udacity\Self-driving_car\projects\Term3\P3\models\research\bstld\tf_object_detection\to_tfrecords.py", line 124, in split_train_labels
train_videos = list(map(lambda x: x['path'].split('/')[-2], train_labels))
File "D:\Udacity\Self-driving_car\projects\Term3\P3\models\research\bstld\tf_object_detection\to_tfrecords.py", line 124, in
train_videos = list(map(lambda x: x['path'].split('/')[-2], train_labels))
IndexError: list index out of range

I am using the updated label files in the label_files folder. Is there any way to solve this? Many thanks~

TF models?

Hi, many thanks for putting this dataset together and also for providing detailed benchmarks and scripts. Any chance we can obtain your saved models as well (NAS-A, SSD)? - thx!

Which Tensorflow version

Hi,

I was unable to figure out what TF version is recommended for use. Could you please clarify? Alternatively, if you could add this (and any other library versions) to requirements.txt please.

Thanks!

Does score_converter should be SOFTMAX?

Hi, you have multiple classes then your score_converter should be SOFTMAX? Why did you use SIGMOID?

get loss_box =0

Hi, When I trained the bosch traffic light dataset with faster rcnn, I got loss_box=0. Does it mean that the label box of the object in the dataset is too small for faster rcnn to detect?

Typo in bosch_to_pascal.py

Line 24:
annotation = ET.Element("annotaion")
should be changed to:
annotation = ET.Element("annotation")

I don't know whether this causes any problems or not.

The dataset can't be used

Hello, Thanks for your great work. However, the dataset that I down load can't be opened. All the images can be open. Is there something wrong with the dataset? Thanks!

OSError

You are getting an error with OSError: [Errno Could not open image path] D:\net\pal-soc1.us.bosch.com\ifs\data\Shared_Exports\deep_learning_data\traffic_lights\university_run1\24068.png

有关数据的标注

请问这个数据集的标注为什么不是整数，我使用的rgb的图像需要针对这一点做什么额外的处理吗？

Focal length

Hi, thanks for sharing your dataset. Could you please provide additional details about the acquisition system? In particular, do you know the value of the focal length of the camera?

Convert to Pascal VOC format

Using a script bosch_to_pascal.py I have xml files without information about labels and bounding boxes

SSD Training

Hi, I tried using the SSD config file to train. However, my training loss is almost constant at 0.346 with continuous spikes in between. After nearly ~150 steps, my loss continuous increases. I have tried tuning the parameters but no change in results unfortunately. My learning rate is also quite low. Is there anything I'm missing or should be looking into?

Thanks!

Cannot Open Dataset Files

Hello,

I downloaded the RGB dataset files from the link provided in the ReadMe. However they were all downloaded as binaries. Looking at the code I see that the input to any of the scripts is a yaml file which I assume would be inside the dataset folders. How can I open/convert to another format?

Thanks!