Closed aiboys closed 2 years ago
Thank you for your interest in our work.
Actually, we pre-define the square 2D Tricube kernel as in Figure 3 (a) and reshape the kernel according to the shape of an object as in Figure 3 (c). You can notice this implementation in here.
In our pre-print version, our work measured the latency using another backbone network (but not included in the paper). With a light version of TricubeNet, which excludes the MAC and Cascade Refinement, our latency is as follows. (Dataset: HRSC2016, GPU: RTX 2080 Ti)
Method | Backbone | Input Size | mAP | FPS |
---|---|---|---|---|
R2 CNN | ResNet-101 | 800x800 | 73.1 | 2 |
RRPN | ResNet-101 | 800x800 | 79.1 | 4 |
RetinaNet-H | ResNet-101 | 800x800 | 82.9 | 14 |
RoI Trans | ResNet-101 | 800x800 | 86.2 | 6 |
RetinaNet-R | ResNet-101 | 800x800 | 89.2 | 10 |
R3Det | ResNet-101 | 800x800 | 89.3 | 12 |
TricubeNet | ResNet-101 | 512x512 | 87.7 | 46 |
TricubeNet (+TTA) | ResNet-101 | 512x512 | 88.6 | 17 |
TricubeNet | DLA-34 | 512x512 | 87.1 | 70 |
TricubeNet (+TTA) | DLA-34 | 512x512 | 88.4 | 30 |
As you can see, the whole inference is very fast with a light version, and there is no bottleneck in the segmentation-based post-processing algorithm.
Hi, Thanks for your open source. It is a good work. But I have the following problems: 1)2d tricube distribution is defined as (1-|x|^3)^r * (1-|y|^3)^r, such distribution is invariant with object's shape (aspect-ratio between height and width). But shown in Fig 7, the output heatmaps are deformable with the shape、aspect ratio and rotation degree. So I amd not sure that there is a modification in your 2d tricube distribution, right?
2) since segmentation based algorithms need post-processing, how about the time-consumption of your TricubeNet. I did not see the comparison experiments with other SOTA algorithms.