Pyramid Scene Parsing Network

Multi-scale features have shown superior performance in detecting objects with different grains.

“Multi-scale features have shown superior performance in detecting objects with different grains [35].”

Ternary quantization and multivariate quantization are harnessed to strike a proper balance between model size and accuracy.

“Hence, ternary quantization and multivariate quantization are harnessed to strike a proper balance between model size and accuracy.”

Virtually all existing architectures for dense prediction are based on convolutional networks.

“Virtually all existing architectures for dense prediction are based on convolutional networks [6,31,34,42,49,50,53].”

Using the feature map up-sampled only once with a down-sampling rate of 16 (P4), FPN reaches mIoU and P.A. of 34.46/76.04, almost identical to the strong baseline reference reported in [16] while only taking about 1/3 of the training time for the same number of iterations.

“Using the feature map up-sampled only once with a down-sampling rate of 16 ( P 4 ), it reaches mIoU and P.A. of 34 . 46 / 76 . 04, almost identical to the strong baseline reference reported in [16] while only taking about 1 / 3 of the training time for the same number of iterations.”

The Pyramid Pooling Module is highly compatible with the FPN architecture by bringing effective global prior representations.

“Empirically we find that the PPM is highly compatible with the FPN architecture by bringing effective global prior representations.”

When trained only on object supervision, UPerNet yields almost identical performance to state-of-the-art PSPNet while requiring only 63% of training time for the same number of epochs.

“When only trained on object supervision, without further enhancements, our framework yields almost identical performance as the state-of-the-art PSPNet, while requiring only 63% of training time for the same number of epochs.”

Batch size is important to generate accurate statistics for tasks like classification, semantic segmentation, and object detection.

“It is worth noting that batch size has proven to be important to generate accurate statistics for tasks like classification [38], semantic segmentation [16] and object detection [39].”

None of the existing remedies (large kernel sizes, atrous convolutions, image/feature pyramids, attention modules, decomposed attention) can fully eliminate the limitation of FCN architecture in receptive field.

“Nonetheless, none of these can fully eliminate the limitation of FCN architecture in receptive field.”

CGNet M3N21's category-wise accuracy is 5.4% less than PSPNet and 5.5% less than DenseASPP.

“while its category-wise accuracy is only 5.4% and 5.5% less respectively.”

CGNet contains 51 convolutional layers with channel numbers 32, 64, and 128, making it a light-weight neural network.

“CGNet only contains 51 layers, and the number of channels in the three stages is 32, 64, 128, respectively.”

CGNet M3N21 is 131 times smaller than PSPNet and 57 times smaller than DenseASPP.

“CGNet M3N21 is 131 and 57 times smaller than PSPNet [44] and DenseASPP [34]”

FPGAs show great improvements in both performance and power consumption in deep learning applications.

“AField-Programmable Gate Array (FPGA) is another viable solution, showing great improvements in both performance and power consumption in deep learning applications.”