Page 54 - My FlipBook
P. 54
工
智
慧
計
畫
Arti cial Intelligence Projects Object detectors equipped with our CSPNet backbone have Search (SM-NAS) model. CSPNet needs a much lower
been shown to achieve superior performance in object image resolution (512x512 vs 800x600) to operate object
detection tasks to other models (see Table 1). It is obvious detectors at 3.35 times the inference speed than SM-NAS
from Table 1 that the inference speed of our CSPNet (67 fps vs 20 fps), yet achieves the same level of accuracy
model is far faster than other models with the same level (42.7% vs 42.8%). This enhanced performance means that
of accuracy requirements (~50-300% faster than existing CSPNet is applicable to broader scenarios, including low-
state-of-the-art models). Compared with more advanced cost video-capture equipment (i.e., any camera or mobile
models, for example, our system is more than 30% faster phone and, thus, is not limited to expensive high-resolution
than the Structural-to-Modular Network Architecture cameras) and non-powerful computing devices.
Table 1 : Performance comparisons between our method and
other state-of-the-art methods.
The advantages of CSPNet become more obvious when the CSPNet can run on a GPU and TX2, achieving excellent
computing resources utilizing it are limited. Compared with performances of 102 fps and 72 fps, respectively. These
other state-of-the-art methods, CSPNet achieves the best performance statistics mean that our CSPNet system can
performance under any requirement of inference speed. extend Arti cial Intelligence of Things (AIoT) to everything
Moreover, relative to the best lightweight model available everywhere. Importantly, since CSPNet reduces memory
today, ThunderNet, the computation speed of CSPNet space requirements by 10-20%, computations by 10-30%,
increases from 133 fps on a GPU to 400 fps, with improved and memory bandwidth by 40-80%, it signi cantly reduces
accuracy from 33.7% to 34.2%. Therefore, CSPNet can development costs for AI-specific ASIC hardware and
use a single GPU to drive 12 camcorders simultaneously subsequent energy loss, while simultaneously improving
and compute traffic flow in real-time. Furthermore, stability.
Figure 3 : Performance of our CSPNet model compared with other state-
of-the-art methods.
52
智
慧
計
畫
Arti cial Intelligence Projects Object detectors equipped with our CSPNet backbone have Search (SM-NAS) model. CSPNet needs a much lower
been shown to achieve superior performance in object image resolution (512x512 vs 800x600) to operate object
detection tasks to other models (see Table 1). It is obvious detectors at 3.35 times the inference speed than SM-NAS
from Table 1 that the inference speed of our CSPNet (67 fps vs 20 fps), yet achieves the same level of accuracy
model is far faster than other models with the same level (42.7% vs 42.8%). This enhanced performance means that
of accuracy requirements (~50-300% faster than existing CSPNet is applicable to broader scenarios, including low-
state-of-the-art models). Compared with more advanced cost video-capture equipment (i.e., any camera or mobile
models, for example, our system is more than 30% faster phone and, thus, is not limited to expensive high-resolution
than the Structural-to-Modular Network Architecture cameras) and non-powerful computing devices.
Table 1 : Performance comparisons between our method and
other state-of-the-art methods.
The advantages of CSPNet become more obvious when the CSPNet can run on a GPU and TX2, achieving excellent
computing resources utilizing it are limited. Compared with performances of 102 fps and 72 fps, respectively. These
other state-of-the-art methods, CSPNet achieves the best performance statistics mean that our CSPNet system can
performance under any requirement of inference speed. extend Arti cial Intelligence of Things (AIoT) to everything
Moreover, relative to the best lightweight model available everywhere. Importantly, since CSPNet reduces memory
today, ThunderNet, the computation speed of CSPNet space requirements by 10-20%, computations by 10-30%,
increases from 133 fps on a GPU to 400 fps, with improved and memory bandwidth by 40-80%, it signi cantly reduces
accuracy from 33.7% to 34.2%. Therefore, CSPNet can development costs for AI-specific ASIC hardware and
use a single GPU to drive 12 camcorders simultaneously subsequent energy loss, while simultaneously improving
and compute traffic flow in real-time. Furthermore, stability.
Figure 3 : Performance of our CSPNet model compared with other state-
of-the-art methods.
52
