MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
2017 · paper · AI and machine learning
Presents MobileNets, small low-latency vision networks built on depthwise separable convolutions, with two hyperparameters for trading accuracy against speed. Experiments show strong ImageNet performance and uses in object detection, fine-grained classification, face attributes and geo-localization.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
7 findings
Substance Table 1 lists "Conv dw / s2 3 × 3 × 1024 dw" with input 7 × 7 × 1024, but the next layer (Conv / s1 1 × 1 × 1024 × 1024) also takes 7 × 7 × 1024, so this layer cannot be stride 2.
- p. 4 Table 1, penultimate conv rows
Conv dw / s2 3 × 3 × 1024 dw 7 × 7 × 1024
Judge's note: I checked the Table 1 crop on page 4. The previous Conv dw / s2 (14×14×512) already reduces the map to 7×7, as the 7×7×512 input of the following Conv / s1 shows. The last dw layer, marked s2, takes 7×7×1024, and the next Conv / s1 and the 7×7 Avg Pool both take 7×7×1024. A stride-2 layer would give about 4×4. The table is internally inconsistent, and the stride should be s1.
Second check: another model agreed
Report a problem with this finding
Substance Table 2 Mult-Adds column (94.86%, 3.06%, 1.19%, 0.18%) sums to 99.29%. The Conv 3×3 share recomputed from Table 1 is about 1.9%, which suggests "1.19%" may be a transposition of 1.90%.
- p. 4 Table 2, Conv 3×3 row
1.19%
Only the authors can say what is right here.
Judge's note: The parameters column sums to 100.00% but the Mult-Adds column does not. The first-layer cost from Table 1 (112×112 output, 3×3×3×32 kernel) is about 10.84M Mult-Adds, which is about 1.9% of 569M; with 1.90% the column sums to 100%. Only the authors can confirm the intended value.
Second check: another model agreed
Report a problem with this finding
Substance Section 4.2 says "Reduced MobileNet is 4% better than AlexNet [19]", but Table 9 gives 60.2% vs 57.2%, a gap of 3.0 points (5.2% relative).
- p. 6 Section 4.2 text on Table 9
MobileNet is 4% better than AlexNet [19]
- p. 6 Table 9
57.2%
- p. 6 Section 4.2, Table 9 discussion
MobileNet is 4% better than AlexNet [19] while being 45×
- p. 6 Section 4.2, paragraph on Table 9
Reduced MobileNet is 4% better than AlexNet [19]
Judge's note: I checked the full page 6 image. Table 9 gives 0.50 MobileNet-160 at 60.2% and AlexNet at 57.2%. Neither the absolute gap nor the relative gap is 4%. The Squeezenet sentence is judged separately under C018.
Second check: another model agreed
Report a problem with this finding
Substance Section 4.2 says "It is also 4% better than Squeezenet [12]", but Table 9 gives 60.2% vs 57.5%, a gap of 2.7 points (4.7% relative).
- p. 6 Section 4.2 text on Table 9
It is also 4% better than Squeezenet [12]
- p. 6 Table 9
57.5%
- p. 6 Section 4.2 text
It is also 4% better than Squeezenet [12] at about the same size
- p. 6 Sec 4.2, text after Table 9
It is also 4% better than Squeezenet
Judge's note: I checked the full page 6 image and Table 9. The absolute gap is 2.7 points, and the adjacent AlexNet sentence uses absolute points. The claimed 4% matches neither the absolute nor the relative reading.
Second check: another model agreed
Report a problem with this finding
Substance Section 4.4 gives the PlaNet MobileNet cost as "0.58 Million mult-adds", which is set against "5.74 billion mult-adds" for the full PlaNet model; the unit should be billion.
- p. 7 Section 4.4
10 million for the final layer and 0.58 Million mult-adds
- p. 7 Section 4.4
has 52 million parameters and 5.74 billion mult-adds
- p. 7 Section 4.4
0.58 Million mult-adds
- p. 7 Sec 4.4
5.74 billion mult-adds. The MobileNet model has only 13 million parameters with the usual 3 million for the body and 10 million for the final layer and 0.58 Million mult-adds
Judge's note: I checked the full page 7 image. A MobileNet body alone costs 569 million Mult-Adds (Table 8), so 0.58 million cannot be right, and the parallel 5.74 billion figure points to 0.58 billion.
Second check: another model agreed
Report a problem with this finding
Substance Section 4.5 describes the baseline face attribute classifier as having "75 million parameters and 1600 million Mult-Adds", but the Table 12 Baseline row lists 1600 Mult-Adds and 7.5 million parameters.
- p. 7 Section 4.5
face attribute classifier with 75 million parameters and 1600 million Mult-Adds
- p. 7 Table 12 Baseline row
Baseline 86.9% 1600 7.5
- p. 7 Section 4.5
a large face attribute classifier with 75 million parameters and 1600 million Mult-Adds
- p. 7 Section 4.5
large face attribute classifier with 75 million parameters and 1600 million Mult-Adds
Only the authors can say what is right here.
Judge's note: I checked the full page 7 image. The text and the Table 12 Baseline row describe the same model with the same 1600M Mult-Adds, but the parameter counts differ by a factor of 10. Only the authors can say which is correct. Table 14 lists FaceNet at 7.5M as well.
Second check: another model agreed
Report a problem with this finding
Reference or label In Table 14 two rows share the label "0.75 MobileNet-128" but report different values (75.2%/166/3.4 and 72.5%/108/3.8), so one label must be wrong.
- p. 8 Table 14, last two rows
0.75 MobileNet-128 75.2% 166 3.4
- p. 8 Table 14, last row
0.75 MobileNet-128 72.5% 108 3.8
Only the authors can say what is right here.
Judge's note: I checked the full page 8 image. Two rows have the same label but different Mult-Adds, and Mult-Adds are fixed by architecture and resolution. One label is wrong, but the files do not show what the intended configuration is.
Second check: another model agreed
Report a problem with this finding
16 minor (typos and the like; most checked by the judge only)
Wording In Table 2 the Mult-Adds column (94.86%, 3.06%, 1.19%, 0.18%) sums to 99.29%, while the Parameters column sums to 100.00%. The remaining 0.71% is not explained.
- p. 4 Table 2
Conv 1 × 1 94.86% 74.59%
Judge's note: I checked the Table 2 crop. The values are as quoted. Layers that are not listed (avg pool, softmax, batchnorm) could plausibly take up the remainder, and the table does not say it is exhaustive. I therefore treat this as a minor observation and not an error.
Second check: another model agreed
Report a problem with this finding
Typo Missing word: "experiments on ImageNet as well a variety" should read "as well as a variety".
- p. 1 Introduction
experiments on ImageNet as well a variety
Judge's note: Crop of page 1 shows the text as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Missing space before citation: "Factorized Networks[34]".
- p. 1 Section 2
Factorized Networks[34]
Judge's note: The page image shows no space before [34], unlike the other citations.
Second check: none (judge only)
Report a problem with this finding
Typo Kernel size is written "Dk × Dk" with a lowercase subscript, whereas D_K is used everywhere else.
- p. 2 Text after Eq. 2
the kernel size Dk × Dk
Judge's note: The crop shows a lowercase k subscript; Eq. 2 directly above uses an uppercase K.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.1: "a 1 × 1 convolution to combine the outputs the depthwise convolution" is missing "of".
- p. 2 Section 3.1, first paragraph
combine the outputs the depthwise convolution
Judge's note: I checked the page 2 crop. The sentence reads exactly as quoted, and "of" is missing.
Second check: none (judge only)
Report a problem with this finding
Typo Missing comma in the list: "the number of output channels N the kernel size Dk × Dk".
- p. 2 Section 3.1, after Eq. 2
the number of output channels N the kernel size Dk × Dk and the feature map
Judge's note: The crop confirms there is no comma after N.
Second check: none (judge only)
Report a problem with this finding
Typo Subject-verb disagreement: "Depthwise separable convolution are made up of two layers".
- p. 3 Section 3.1, second paragraph
Depthwise separable convolution are made up of two
Judge's note: Visible in the crop of page 3.
Second check: none (judge only)
Report a problem with this finding
Typo Word errors in Section 3.2: "since their are so few parameters" (there), "95% of it's computation time" (its), "reduce the amount image of distortions" (amount of image distortions).
- p. 4 Section 3.2
since their are so few parameters in them
- p. 4 Section 3.2
MobileNet spends 95% of it’s computation time
- p. 4 Section 3.2
reduce the amount image of distortions
- p. 4 Section 3.2
since their are so few parameters
Judge's note: All three appear in the page 4 crops as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo "only reduces accuracy by 1% on ImageNet was saving tremendously on mult-adds" should read "while saving".
- p. 5 Section 4.1
accuracy by 1% on ImageNet was saving tremendously on
Judge's note: Visible in the crop of page 5.
Second check: none (judge only)
Report a problem with this finding
Typo The Section 4.4 heading is misspelled: "4.4. Large Scale Geolocalizaton" (should be "Geolocalization").
- p. 6 Section 4.4 heading
Large Scale Geolocalizaton
- p. 6 Section 4.4 heading
4.4. Large Scale Geolocalizaton
Judge's note: I checked the full page 6 image. The heading reads "Geolocalizaton" and is missing an "i".
Second check: none (judge only)
Report a problem with this finding
Typo The Figure 6 caption reads "Example objection detection results"; it should read "object detection".
- p. 7 Figure 6 caption
Example objection detection results
- p. 7 Figure 6 caption
Example objection detection results using MobileNet
Judge's note: The wording is visible in the caption in the crop.
Second check: none (judge only)
Report a problem with this finding
Typo Section 4.5 reads "as the in-house while consuming only 1% the Multi-Adds": "Multi-Adds" should be "Mult-Adds" as elsewhere, "of" is missing, and "in-house" has no noun.
- p. 7 Section 4.5
consuming only 1% the Multi-Adds
- p. 7 Section 4.5
in-house while consuming only 1% the Multi-Adds
- p. 7 Sec 4.5
while consuming only 1% the Multi-Adds
- p. 7 Table 12
1.0 MobileNet-224 88.7% 568
Judge's note: I checked the full page 7 image, and the wording is as quoted. I treat the 1% claim as a loose reference to the smallest model (15/1600 is about 0.94%, at 86.4% vs 86.9% mAP, which the text calls "similar"), and I do not count it as a separate error.
Second check: none (judge only)
Report a problem with this finding
Typo Table 13, last row: "Mobilenet 19.8% 30.5 6.1"; the other rows of the same table write "MobileNet".
- p. 7 Table 13, Faster-RCNN 600 block
Mobilenet 19.8%
- p. 7 Table 13, last model row
Mobilenet 19.8% 30.5 6.1
Judge's note: Visible in the Table 13 crop. Cosmetic inconsistency within one table.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [12] title prints "50x fewer parameters and¡ 1mb model size"; it should read "and <1MB model size".
- p. 8 Reference [12]
50x fewer parameters and¡ 1mb model size
Judge's note: The '¡' glyph is visible in the rendered PDF (crop C029-a), so this is not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo The Conclusion reads "releasing models in Tensor Flow"; elsewhere the paper writes "TensorFlow".
- p. 8 Conclusion
releasing models in Tensor Flow
Judge's note: Visible in the Conclusion on page 8; Section 3.2 writes "TensorFlow".
Second check: none (judge only)
Report a problem with this finding
Typo Reference [26] reads "L. Sifre. Rigid-motion scattering for image classification. PhD thesis, Ph. D. thesis, 2014." The words "PhD thesis" are duplicated, and the entry gives no awarding institution.
- p. 9 Reference [26]
PhD thesis, Ph. D. thesis, 2014.
Judge's note: I checked the crop and page 9 of the PDF. The duplicated "PhD thesis, Ph. D. thesis" is printed as quoted, so this is a cosmetic bibliographic error.
Second check: none (judge only)