We will continue to update the data and code corresponding to the paper on multi-path dynamic tracking strategy for MVR-assisted small object tracking!
Hardware environment for this technical experiment:
Intel (R) Xeon (R) Silver 4215R CPU @ 3.20 GHz. We tested the inference speed at FP32 precision on our hardware. Specifically, our complete hardware-related metrics are: NVIDIA GeForce RTX 3090 model, CUDA cores of 8.6, total video memory of 24GB, and matrix multiplication performance of 21.31~22.04 TFLOPS.
The required environment is Python 3.8 (all packages in the code are installed based on Python 3.8).
our tracking data: https://pan.baidu.com/s/10X-zheumV1OMetYj8IV40Q?pwd=8888
Follow Table: The achievements of our MVR (TFFN Module Assisted small object detecting)
| Network | TFFN (✔:used) | REC⬆ | PRE〰 | FPR〰 | FNR⬇ | F1-score⬆ | mAP@0.5:0.95⬆ | mAP@0.75⬆ | mAP@0.5⬆ |
|---|---|---|---|---|---|---|---|---|---|
| Deformable DETR | 9.79% | 86.61% | 1.47% | 90.21% | 17.59% | 24.30% | 26.12% | 41.69% | |
| ✔ | 20.65%⬆ | 77.43% | 0.68% | 79.35%⬇ | 32.60%⬆ | 41.85%⬆ | 42.89%⬆ | 76.11%⬆ | |
| FasterNet | 7.96% | 51.93% | 1.61% | 92.04% | 13.80% | 15.30% | 12.63% | 32.31% | |
| ✔ | 13.26%⬆ | 42.58% | 1.92% | 86.74%⬇ | 20.23%⬆ | 18.40%⬆ | 14.60%⬆ | 40.23%⬆ | |
| FasterRCNN | 24.42% | 84.64% | 1.66% | 75.58% | 37.90% | 51.38% | 60.82% | 76.42% | |
| ✔ | 31.80%⬆ | 87.77% | 0.83% | 68.20%⬇ | 46.69%⬆ | 58.45%⬆ | 68.26%⬆ | 88.86%⬆ | |
| MobileNetV3 | 23.31% | 85.22% | 1.99% | 76.69% | 36.60% | 47.73% | 55.72% | 73.45% | |
| ✔ | 31.02%⬆ | 86.10% | 1.01% | 68.98%⬇ | 45.61%⬆ | 54.74%⬆ | 62.36%⬆ | 86.77%⬆ | |
| MobileNetV4 | 6.97% | 57.74% | 1.17% | 93.03% | 12.44% | 17.90% | 15.04% | 37.30% | |
| ✔ | 10.59%⬆ | 47.56% | 1.27% | 89.41%⬇ | 17.32%⬆ | 20.36%⬆ | 15.99%⬆ | 44.67%⬆ | |
| ResNet152 | 18.47% | 88.81% | 1.59% | 81.53% | 30.59% | 46.10% | 53.55% | 71.35% | |
| ✔ | 27.92%⬆ | 87.16% | 0.76% | 72.08%⬇ | 42.29%⬆ | 55.11%⬆ | 62.66%⬆ | 87.41%⬆ | |
| YOLOv8 | 24.17% | 81.71% | 2.06% | 75.83% | 37.30% | 55.36% | 66.98% | 78.54% | |
| ✔ | 29.72%⬆ | 88.71% | 0.95% | 70.28%⬇ | 44.53%⬆ | 61.55%⬆ | 72.93%⬆ | 89.85%⬆ |
With the assistance of our PoseSet, the drone can selectively choose multiple optimal tracking poses, while avoiding target loss due to the model's poor detection perspective.
Follow Videos: The Results of our MVR (MPST Module Assisted small object Tracking)
Bad Pose (bad Tracking!!!):
PoseSet1 (good Tracking):
PoseSet2:
PoseSet3:
PoseSet4:
