Here is our pytorch implementation of the EfficientDet model to Detection and classification of small objects
Our project running on Colab pro
Here is our pytorch implementation of the model described in the Paper EfficientDet: Scalable and Efficient Object Detection paper (Note: We also provide pre-trained weights, which you could see in Drive at folder "train_weights"
What You Will Learn
- How to load your custom image detection data from Roboflow
- How to instatiate a pytorch EfficientDet model
- How to train the EfficientDet model
- How to use the model for quick inference
- How to export the model weights for future inference
- How to reload the model weights
- How to evaluate mAP
Datasets Drive
| Dataset | #Images | Image size (AVG) | #Object | Object per image (AVG) | Average object area ratio (%) |
|---|---|---|---|---|---|
| Chess | 289 | 2048x1271 | 2870 | 9.9 | 1% |
| MAV-VID | 18952 | 1920x1080 | 20288 | 1.1 | 1% |
| Full-UAV | 2732 | 1100x900 | 3228 | 1.2 | 1% |
Experiments
Our Datasets divided as shown in Table below
| Datasets | Train set(#Image) | Train set(#Object) | Valid set(#Image) | Valid set(#Object) | Test set(#Image) | Test set(#Object) | Sum(#Image) | Sum(#Object) |
|---|---|---|---|---|---|---|---|---|
| MAV-VID | 12517 | 13842 | 4713 | 4717 | 1722 | 1729 | 18952 | 20288 |
| Full-UAV | 1959 | 2426 | 502 | 514 | 271 | 288 | 2732 | 3228 |
| MAV-VID + Full-UAV | 14476 | 16268 | 5215 | 5231 | 1993 | 2017 | 21684 | 23516 |
| Chess | 202 | 2108 | 58 | 286 | 29 | 376 | 289 | 2870 |
We trained our model by using Colab Pro
Chess Experiments
| Augmentation | Weight_download | Epochs (stopped early) | Training Time(sec per epoch) | mAP |
|---|---|---|---|---|
| Baseline | Weight | 48 | 25 | 96.71% |
| Flip Horizontal | Weight | 45 | 32 | 95.95% |
| Flip Vertical | Weight | 69 | 29 | 94.41% |
| Flip Horizontal or Vertical | Weight | 63 | 29 | 97.35% |
| Flip Horizontal and Vertical | Weight | 97 | 25 | 95.16% |
| Hue | Weight | 65 | 27 | 93.79% |
| Saturation | Weight | 95 | 25 | 95.3% |
| Brightness | Weight | 65 | 27 | 89.62% |
| Saturation and Brightness | Weight | 55 | 28 | 95.94% |
| Saturation or Brightness | Weight | 53 | 28 | 89.46% |
| Grayscale | Weight | 35 | 25 | 95.82% |
Full-UAV Experiments
| Augmentation | Weight_download | Epochs (stopped early) | Training Time(sec per epoch) | mAP |
|---|---|---|---|---|
| Baseline | Weight | 6 | 660 | 74.26% |
| Flip Horizontal | Weight | 9 | 293 | 80.0% |
| Hue | Weight | 8 | 225 | 48.94% |
| Saturation | Weight | 18 | 180 | 62.0% |
| Brightness | Weight | 23 | 153 | 63.74% |
| Saturation and Brightness | Weight | 14 | 184 | 64.22% |
| Saturation or Brightness | Weight | 12 | 195 | 59.47% |
| Grayscale | Weight | 6 | 280 | 68.29% |
MAV-VID Experiments
| Augmentation | Weight_download | Epochs (stopped early) | Training Time(sec per epoch) | mAP |
|---|---|---|---|---|
| Baseline | Weight | 3 | 6640 | 73.14% |
| Flip Horizontal | Weight | 4 | 3900 | 82.0% |
| Saturation | Weight | 8 | 2925 | 79.63% |
| Brightness | Weight | 8 | 2400 | 76.38% |
| Grayscale | Weight | 4 | 3900 | 77.15% |
Combine MAV-VID + Full-UAV Experiments
| Augmentation | Weight_download | Epochs (stopped early) | Training Time(sec per epoch) | mAP |
|---|---|---|---|---|
| Baseline | Weight | 3 | 7220 | 77.38% |
| Flip Horizontal | Weight | 7 | 377 | 83.25% |
| Saturation | Weight | 7 | 3394 | 79.04% |
| Brightness | Weight | 7 | 3745 | 80.65% |
| Grayscale | Weight | 7 | 3514 | 78.19% |
| Saturation or Brightness | Weight | 6 | 3590 | 82.57% |
Results
Some predictions are shown below:
References
Appreciate the great work from the following repositories: