Introduction
Safety Gear is a public domain dataset and imported in deepview format for Au-Zone Technologies. This dataset has a high unbalance factor among classes. Original dataset contains 4 different classes Vest, HardHat, Boots and Mask. In order to make models more robust during the inference, we removed 2 classes from the dataset (Boots and Mask) and reannotate the entire dataset so it can detect unsafe workers. The new distributions of classes within the dataset is the following: HardHat, No_HardHat, No_Vest and Vest. By doing this new refinement, we can guarantee a better confidence for all the models trained over this dataset.
Use Case
When performing hazardous tasks, workers should wear the correct protection all the time. With this dataset we can build a solution capable of detecting which worker is not wearing the right protection. This is useful to prevent and monitoring undesired accidents at work.
Dataset Stats
This dataset contains more than 2700 annotated images describing different scenarios where protection gear is mandatory to be used. Among the images, 2700 are designated as training samples while more than 300 are used as test set.
HardHat | No HardHat | No Vest | Vest | |
Train | 2203 | 340 | 571 | 1707 |
Test | 242 | 29 | 32 | 199 |
Table 1: Dataset distribution samples (Train images vs Test images)
Dataset Samples
The following are random samples from the dataset.
Evaluation Results
Table 2: Evaluation metrics at (mAP @ 0.5). The following results were obtained after training ModelPack over 30 epochs.
Model | Float RTM | Per-Channel-Quantized RTM | Per-Tensor-Quantized RTM | Input Resolution |
ModelPack | 53.65 % | 52.67 % | 52.38 % | 416x416 |
ModelPack | 58.13 % | 58.28 % | 57.22 % | 640x640 |
Table 3: NXP i.MX 8M Plus Inference timings on CPU, GPU and NPU (milliseconds) at 416x416 input resolution
Model | Float RTM | Per-Channel-Quantized RTM | Per-Tensor-Quantized RTM |
ModelPack CPU | 491.57 | 407.08 | 419.51 |
ModelPack GPU | 3546.49 | 2924.66 | 1050.58 |
ModelPack NPU | 3828.17 | 18.1 |
15.21 |
Table 4: NXP i.MX 8M Plus Inference timings on CPU, GPU and NPU (milliseconds) at 640x640 input resolution
Model | Float RTM | Per-Channel-Quantized RTM | Per-Tensor-Quantized RTM |
ModelPack CPU | - | 675.69 | 675.88 |
ModelPack GPU | 5942.90 | 5305.40 | 2547.14 |
ModelPack NPU | 8658.90 | 40.40 | 36.12 |
Conclusions
In this article we have introduced our Safety Gear dataset as well as the results our ModelPack achieved after training only epochs.
Download
Download Safety Gear dataset and pre-trained checkpoints/models
Comments
0 comments
Please sign in to leave a comment.