Proposal learning for semi-supervised object detection

Abstract

A method for generating a neural network for detecting one or more objects in images includes generating one or more self-supervised proposal learning losses based on the one or more proposal features and corresponding proposal feature predictions. One or more consistency-based proposal learning losses are generated based on noisy proposal feature predictions and the corresponding proposal predictions without noise. A combined loss is generated using the one or more self-supervised proposal learning losses and one or more consistency-based proposal learning losses. The neural network is updated based on the combined loss.

Claims

1. A method for generating a neural network for detecting one or more objects in images, comprising: generating one or more region proposals that may contain objects for each image of a set of unlabeled images; determining one or more proposal features for each of the region proposals and corresponding proposal feature predictions using a proposal convolutional feature map; adding noise to the proposal convolutional feature map to generate a noisy proposal convolutional feature map; generating one or more noisy proposal features using the noisy proposal convolutional feature map; generating one or more self-supervised proposal learning losses based on the one or more proposal features and corresponding proposal feature predictions, and the one or more noisy proposal features and corresponding noisy proposal feature predictions; generating one or more consistency-based proposal learning losses based on noisy proposal feature predictions and the corresponding proposal predictions without noise; generating a combined loss using the one or more self-supervised proposal learning losses and one or more consistency-based proposal learning losses; and updating the neural network based on the combined loss.

2. The method of claim 1, wherein the determining the one or more self-supervised proposal learning losses includes: generating one or more proposal location losses using the unlabeled images to learn context-aware features; generating one or more contrastive losses using the unlabeled images to learn noise-robust proposal features; and generating a first self-supervised proposal learning loss based on the one or more proposal location losses and one or more contrastive losses.

3. The method of claim 2, wherein the generating the one or more contrastive losses include: generating a first contrastive loss using the noisy proposal features.

4. The method of claim 3, wherein the generating the one or more contrastive losses include: adding noise to the unlabeled images or intermediate features to generate noisy unlabeled images or noisy intermediate features respectively; and generating a second contrastive loss using at least one of the noisy unlabeled images and noisy intermediate features.

5. The method of claim 1, wherein the generating the one or more consistency-based proposal learning losses includes: generating a first consistency loss using the unlabeled images for bounding box classification predictions; generating a second consistency loss using the unlabeled images for bounding box regression predictions; and generating the consistency-based proposal learning loss based on the first consistency loss and the second consistency loss.

6. The method of claim 1, further comprising: determining one or more fully-supervised losses of the neural network using a set of labeled images; and generating the combined loss using the one or more self-supervised proposal learning losses, the one or more consistency-based proposal learning losses, and the one or more fully-supervised losses.

7. The method of claim 6, wherein each of the labeled images includes at least one of an image-level class label and a bounding box label.

8. A non-transitory machine-readable medium comprising a plurality of machine-readable instructions which, when executed by one or more processors, are adapted to cause the one or more processors to perform a method for generating a neural network for detecting one or more objects in images comprising: generating one or more region proposals that may contain objects for each image of a set of unlabeled images; determining one or more proposal features for each of the region proposals and corresponding proposal feature predictions using a proposal convolutional feature map; adding noise to the proposal convolutional feature map to generate a noisy proposal convolutional feature map; generating one or more noisy proposal features using the noisy proposal convolutional feature map; generating one or more self-supervised proposal learning losses based on the one or more proposal features and corresponding proposal feature predictions and the one or more noisy proposal features and corresponding noisy proposal feature predictions; generating one or more consistency-based proposal learning losses based on noisy proposal feature predictions and the corresponding proposal predictions without noise; generating a combined loss using the one or more self-supervised proposal learning losses and one or more consistency-based proposal learning losses; and updating the neural network based on the combined loss.

9. The non-transitory machine-readable medium of claim 8, wherein the determining the one or more self-supervised proposal learning losses includes: generating one or more proposal location losses using the unlabeled images to learn context-aware features; generating one or more contrastive losses using the unlabeled images to learn noise-robust proposal features; and generating a first self-supervised proposal learning loss based on the one or more proposal location losses and one or more contrastive losses.

10. The non-transitory machine-readable medium of claim 9, wherein the generating the one or more contrastive losses include: generating a first contrastive loss using the noisy proposal features.

11. The non-transitory machine-readable medium of claim 10, wherein the generating the one or more contrastive losses include: adding noise to the unlabeled images or intermediate features to generate noisy unlabeled images or noisy intermediate features respectively; and generating a second contrastive loss using at least one of the noisy unlabeled images and noisy intermediate features.

12. The non-transitory machine-readable medium of claim 8, wherein the generating the one or more consistency-based proposal learning losses includes: generating a first consistency loss using the unlabeled images for bounding box classification predictions; generating a second consistency loss using the unlabeled images for bounding box regression predictions; and generating the consistency-based proposal learning loss based on the first consistency loss and the second consistency loss.

13. The non-transitory machine-readable medium of claim 8, wherein the method further comprises: determining one or more fully-supervised losses of the neural network using a set of labeled images; and generating the combined loss using the one or more self-supervised proposal learning losses, the one or more consistency-based proposal learning losses, and the one or more fully-supervised losses.

14. The non-transitory machine-readable medium of claim 13, wherein each of the labeled images includes at least one of an image-level class label and a bounding box label.

15. A system, comprising: a non-transitory memory; and one or more hardware processors coupled to the non-transitory memory and configured to read instructions from the non-transitory memory to cause the system to perform a method for generating a neural network for detecting one or more objects in images comprising: generating one or more region proposals that may contain objects for each image of a set of unlabeled images; determining one or more proposal features for each of the region proposals and corresponding proposal feature predictions using a proposal convolutional feature map; adding noise to the proposal convolutional feature map to generate a noisy proposal convolutional feature map; generating one or more noisy proposal features using the noisy proposal convolutional feature map; generating one or more self-supervised proposal learning losses based on the one or more proposal features and corresponding proposal feature predictions and the one or more noisy proposal features and corresponding noisy proposal feature predictions; generating one or more consistency-based proposal learning losses based on noisy proposal feature predictions and the corresponding proposal predictions without noise; generating a combined loss using the one or more self-supervised proposal learning losses and one or more consistency-based proposal learning losses; and updating the neural network based on the combined loss.

16. The system of claim 15, wherein the determining the one or more self-supervised proposal learning losses includes: generating one or more proposal location losses using the unlabeled images to learn context-aware features; generating one or more contrastive losses using the unlabeled images to learn noise-robust proposal features; and generating a first self-supervised proposal learning loss based on the one or more proposal location losses and one or more contrastive losses.

17. The system of claim 15, wherein the generating the one or more contrastive losses include: generating a first contrastive loss using the noisy proposal features.

18. The system of claim 15, wherein the generating the one or more contrastive losses include: adding noise to the unlabeled images or intermediate features to generate noisy unlabeled images or noisy intermediate features respectively; and generating a second contrastive loss using at least one of the noisy unlabeled images and noisy intermediate features.

19. The system of claim 15, wherein the generating the one or more consistency-based proposal learning losses includes: generating a first consistency loss using the unlabeled images for bounding box classification predictions; generating a second consistency loss using the unlabeled images for bounding box regression predictions; and generating the consistency-based proposal learning loss based on the first consistency loss and the second consistency loss.

20. The system of claim 15, wherein the method further comprises: determining one or more fully-supervised losses of the neural network using a set of labeled images; and generating the combined loss using the one or more self-supervised proposal learning losses, the one or more consistency-based proposal learning losses, and the one or more fully-supervised losses.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 is a simplified diagram of a computing device according to some embodiments.

(2) FIG. 2 is a simplified diagram of a method of training a neural network model according to some embodiments.

(3) FIG. 3 is a simplified diagram of a neural network model according to some embodiments.

(4) FIG. 4 is a simplified diagram of a trained neural network model for interference according to some embodiments.

(5) In the figures, elements having the same designations have the same or similar functions.

DETAILED DESCRIPTION

(6) FIG. 1 is a simplified diagram of a computing device 100 according to some embodiments. As shown in FIG. 1, computing device 100 includes a processor 110 coupled to memory 120. Operation of computing device 100 is controlled by processor 110. And although computing device 100 is shown with only one processor 110, it is understood that processor 110 may be representative of one or more central processing units, multi-core processors, microprocessors, microcontrollers, digital signal processors, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), graphics processing units (GPUs) and/or the like in computing device 100. Computing device 100 may be implemented as a stand-alone subsystem, as a board added to a computing device, and/or as a virtual machine.

(7) Memory 120 may be used to store software executed by computing device 100 and/or one or more data structures used during operation of computing device 100. Memory 120 may include one or more types of machine readable media. Some common forms of machine readable media may include floppy disk, flexible disk, hard disk, magnetic tape, any other magnetic medium, CD-ROM, any other optical medium, punch cards, paper tape, any other physical medium with patterns of holes, RAM, PROM, EPROM, FLASH-EPROM, any other memory chip or cartridge, and/or any other medium from which a processor or computer is adapted to read.

(8) Processor 110 and/or memory 120 may be arranged in any suitable physical arrangement. In some embodiments, processor 110 and/or memory 120 may be implemented on a same board, in a same package (e.g., system-in-package), on a same chip (e.g., system-on-chip), and/or the like. In some embodiments, processor 110 and/or memory 120 may include distributed, virtualized, and/or containerized computing resources. Consistent with such embodiments, processor 110 and/or memory 120 may be located in one or more data centers and/or cloud computing facilities.

(9) As shown, memory 120 includes a neural network module 130 that may be used to implement and/or emulate the neural network systems and models described further herein and/or to implement any of the methods described further herein. In some examples, neural network module 130 may be used to translate structured text. In some examples, neural network module 130 may also handle the iterative training and/or evaluation of a translation system or model used to translate the structured text. In some examples, memory 120 may include non-transitory, tangible, machine readable media that includes executable code that when run by one or more processors (e.g., processor 110) may cause the one or more processors to perform the counting methods described in further detail herein. In some examples, neural network module 130 may be implemented using hardware, software, and/or a combination of hardware and software. As shown, computing device 100 receives input 140, which is provided to neural network module 130, neural network module 130 then generates output 150.

(10) FIG. 2 is a simplified diagram of a method 200 for training a neural network model using proposal learning to learn proposal features and predictions from both labeled and unlabeled data and performing a task using the trained neural network model according to some embodiments. One or more of the processes 202-214 of method 200 may be implemented, at least in part, in the form of executable code stored on non-transitory, tangible, machine-readable media that when run by one or more processors may cause the one or more processors to perform one or more of the processes 202-214. In some embodiments, the method 200 may correspond to the method used by neural network module 130 to perform training and/or perform inference using the trained neural network model for various tasks (e.g., classification, prediction, diagnoses and prediction on medical diseases, image recognition, natural language processing, etc.).

(11) The method 200 performs proposal learning to learn proposal features and predictions from both labeled and unlabeled data. The method 200 includes self-supervised proposal earning and consistency-based proposal learning. During the self-supervised proposal learning, a proposal location loss and a contrastive loss are used to learn context-aware and noise-robust proposal features respectively. During the consistency-based proposal learning, consistency losses are applied to both bounding box classification and regression predictions of proposals to learn noise-robust proposal features and predictions.

(12) The method 200 begins at block 202, where a processor performs initialization for training a neural network model. At block 202, the method 200 may receive various inputs. For example, the inputs may include a set of labeled data and a set of unlabeled data.

(13) The method 200 may proceed to block 204, where the processor may perform supervised training using the set of labeled data. In some examples, a fully-supervised loss is generated for each labeled data in the set of labeled data.

(14) The method 200 may proceed to blocks 206 and 208, where the process may perform training using the unlabeled training data. In some examples, at block 206, a self-supervised proposal learning loss is generated for each unlabeled data in the set of unlabeled data. In some examples, at block 208, a consistency-based proposal learning loss is generated for each unlabeled data in the set of unlabeled data.

(15) The method 200 may proceed to block 210, where the processor may generate a combined loss using two or more of the fully-supervised losses, the self-supervised proposal learning losses, and the consistency-based proposal learning losses.

(16) The method 200 may proceed to block 212, where the processor may train the neural network (e.g., including object detectors) by performing optimization using the combined loss. Various optimization methods, e.g., back propagation, may be used.

(17) The method 200 may proceed to block 214, where the trained network is deployed to perform an inference process for a particular task (e.g., classification, prediction, diagnoses and prediction on medical diseases, image recognition, natural language processing, etc.).

(18) Referring to the example of FIG. 3, illustrated therein is a neural network system 300 for object detection. As shown in FIG. 3, the neural network system 300 is trained by performing proposal learning to learn proposal features and predictions from both labeled and unlabeled data (e.g., training steps 202-212 of method 200 of FIG. 2). In the system of FIG. 3, all the steps 352 through 396 have forward computations during the training, and steps with solid lines (e.g., steps 354, 356, 360, 362, 364, 366, 368, 370, 372, 374, 376, 378, 380, 382, 392, 394, 396) have back-propagation computations as well. In other words, steps with dashed lines (e.g., steps 352, 358, 390, and 398) do not have back-propagation computations. Also, steps 392 and 394 are performed only with labeled data, with ground label truth G 390. The other steps may be performed with unlabeled data, labeled data, or a combination thereof.

(19) Basic Components

(20) The neural network 300 includes a number of neural network components, each may include a neural network. The neural network components include for example, a backbone neural network 304, a Region Proposal Network (RPN) 360, a region of interest align (RoIAlign) layer 310, a Region-based CNN (R-CNN) 314, and a Self-Supervised Proposal Learning (SSPL) 338.

(21) In some embodiments, during a training process, at step 352, an image 302 (denoted as I) from a training set is fed into the backbone neural network 304 with parameters θ.sup.b. The backbone neural network 304 may include a convolutional neural network (CNN), for example, a ResNet-50 model with feature pyramid networks), or any other suitable neural networks. At step 354, the backbone neural network 304 generates image convolutional feature maps 306, denoted as F.sup.b(I; θ.sup.b).

(22) In some embodiments, at step 356, the RPN 360 with parameters θ.sup.rpn receives image convolutional feature maps 306 F.sup.b(I; θ.sup.b) (also referred to as F.sup.b) as inputs. The RPN 360, at step 358, generates region proposals 308 (denoted as P(F.sup.b(I; θ.sup.b); θ.sup.rpm), also referred to as P), based on the image convolutional feature maps 306 F.sup.b(I; θ.sup.b).

(23) In some embodiments, at step 360, the RoIAlign layer 310 takes each proposal P.sub.n=(x.sub.n, y.sub.n, w.sub.n, h.sub.n) in region proposals 308 P and F.sup.b as inputs, where (x.sub.n, y.sub.n, w.sub.n, h.sub.n) denotes the location of the i.sup.th proposal P.sub.n, i∈{1, 2, . . . , N}, and N is the number of proposals in P.

(24) At step 362, the RoIAlign layer 310 extracts proposal convolutional feature maps 312 F.sub.n.sup.p-conv, which is a simplification of F.sup.p-conv(F.sup.b, P.sub.n) by dropping the dependence on F.sup.b, P.sub.n.

(25) In some embodiments, at step 364, proposal convolutional feature maps 312 F.sub.n.sup.p-conv is fed into a Region-based CNN (R-CNN) 314.

(26) In some embodiments, at step 366, the R-CNN 314 generates proposal features 316 F.sup.p(F.sub.n.sup.p-conv; θ.sup.r-cnn), also referred to as F.sub.n.sup.p. θ.sup.r-cnn denotes parameters of the R-CNN 314 to generate proposal features 316. At step 368, the R-CNN 314 generates predictions 320. The predictions 320 may include different types of predictions. For example, the predictions 320 may include bounding box classification predictions 322 C.sup.p(F.sup.p(F.sub.n.sup.p-conv; θ.sup.r-cnn); θ.sup.cls), also referred to as C.sub.n.sup.p, where θ.sup.cls denotes parameters of the R-CNN to generate bounding box classification predictions. For further example, the predictions 320 may include bounding box regression predictions 324 R.sup.p(F.sup.p(F.sub.n.sup.p-conv; θ.sup.r-cnn; θ.sup.reg)), also referred to as R.sub.n.sup.p, where θ.sup.reg denote parameters of the R-CNN to generate bounding box regression predictions.

(27) Fully-Supervised Learning

(28) As shown in FIG. 3, a supervised loss is generated using labeled data (e.g., at block 204 of method 200). The supervised loss may be generated using steps 392 and 394 are performed only for labeled data having ground truth label G, and are not performed for unlabeled data. Specifically, for each input image 302 that is a labeled image (also referred to as labeled data), the neural network 300 is trained based on a supervised loss L.sup.sup, e.g., as follows:

(29) $\begin{array}{cc}{L}^{\sup }\left(I,𝒢;{\theta }^b,{\theta }^{\mathrm{rpn}},{\theta }^{r-\mathrm{cnn}},{\theta }^{\mathrm{cls}},{\theta }^{\mathrm{reg}}\right)=L^{\mathrm{rpn}}\left(I,𝒢;{\theta }^b,{\theta }^{\mathrm{rpn}}\right)+\frac{1}{N}\underset{n}{.Math.}{L}^{r-\mathrm{cnn}}\left(I,P_n,𝒢,{\theta }^b,{\theta }^{r-\mathrm{cnn}},{\theta }^{\mathrm{cls}},{\theta }^{\mathrm{reg}}\right)=L^{\mathrm{rpn}}\left(F^b,𝒢;{\theta }^{\mathrm{rpn}}\right)+\frac{1}{N}.Math.L^{r-\mathrm{cnn}}\left(C_n^p,R_n^p,𝒢\right).& \left(1\right)\end{array}$

(30) Specifically, the first term L.sup.rpn and second term L.sup.r-cnn denote the RPN loss and R-CNN loss respectively, G is the ground truth labels, and this supervised loss L.sup.sup is optimized with regard to θ.sup.b, θ.sup.rpn, θ.sup.r-cnn, θ.sup.cls, θ.sup.reg to train the neural network 300 (also referred to as object detectors 300) during the back-propagation process (e.g., steps with solid lines as shown in FIG. 3). For example, as shown in FIG. 3, at step 392, L.sup.rpn 326 for labeled data I with ground label truth G is generated using RPN 360 with respect to parameters θ.sup.rpn and θ.sup.b. For further, as shown in FIG. 3, at step 394, L.sup.r-cnn 328 for labeled data I with ground label truth G is generated using R-CNN 314 with regard to θ.sup.b, θ.sup.rpn, θ.sup.r-cnn, θ.sup.cls, θ.sup.reg.

(31) Proposal Learning

(32) For each input image 302 that is an unlabeled image, there is no ground truth labels G. To train the neural network 300 with an unlabeled image, a proposal learning (e.g., at blocks 206 and 208 of method 200) is performed. Such a proposal learning includes using a self-supervised proposal learning module to learn proposal features (e.g., F″) from unlabeled data (e.g., at block 206 of method 200), and a consistency-based proposal learning module to learn predictions (e.g., bounding box classification predictions 322 C.sub.n.sup.p, bounding box regression predictions 324 R.sub.n.sup.p, any other suitable predictions, and/or a combination thereof) (e.g., at block 208 of method 200).

(33) It is noted that while the discussion below focuses on optimizing the R-CNN 314 using the unlabeled data, it is possible to also benefit RPN 360 from the unlabeled data. The discussion below focuses on optimizing R-CNN-related parts including R-CNN 314 using the unlabeled data, because 1) the final object detection results are from R-CNN-related parts and thus improving the R-CNN-related parts will benefit object detectors directly; 2) gradients will also be back-propagated from the R-CNN-related parts to the backbone 304 to learn better image convolutional feature maps 306, which may benefit RPN 360.

(34) In the example of FIG. 3, at step 370, noise (e.g., random noise {ϵ.sub.nk}.sub.k=1.sup.K or any other suitable noise) is added to proposal convolutional feature maps 312 F.sub.n.sup.p-conv, and generates a set of noisy proposal convolutional feature maps 330 {circumflex over (F)}.sub.n.sup.p-conv, e.g.,
.sub.n.sup.p-conv={{circumflex over (F)}.sup.p-conv(F.sub.n.sup.p-conv,ϵ.sub.nk)}.sub.k=1.sup.K.

(35) At step 366, the R-CNN 314 generates proposal features 316 F.sup.p(F.sub.n.sup.p-conv; θ.sup.r-cnn), also referred to as F.sub.n.sup.p. θ.sup.r-cnn denotes parameters of the R-CNN 314 to generate proposal features 316.

(36) Similar to steps 366 and 368 to generate proposal features 316 F.sup.p and predictions 320, at steps 374 and 376 are performed to generate noisy proposal features 318 and noisy predictions 324 and 326. Specifically, at step 374, the R-CNN 314 generates noisy proposal features 318 .sub.n.sup.p={{circumflex over (F)}.sub.nk.sup.p}.sub.k=1.sup.K using noisy proposal convolutional feature maps 330. At step 376, R-CNN 314 generates noisy predictions 332. The noisy predictions 332 may include different types of noisy predictions. For example, the noisy predictions 332 may include noisy bounding box classification predictions 334, e.g., .sub.n.sup.p={Ĉ.sub.nk.sup.p}.sub.k=1.sup.K and noisy bounding box regression predictions 336 .sub.n.sup.p={{circumflex over (R)}.sub.nk.sup.p}.sub.k=1.sup.K.

(37) Self-Supervised Proposal Learning

(38) In various embodiments, the self-supervised proposal learning module uses different types of losses to learn different types of features. For example, a proposal location loss is used to learn context-aware proposal features, and a contrastive loss is used to learn noise-robust proposal features respectively.

(39) In some embodiments, the self-supervised proposal learning module uses a proposal location loss 340 L.sup.self-loc to learn context-aware proposal features. To compute the proposal location loss 340 L.sup.self-loc, a proposal location prediction as the pretext task is used. At process 378, proposal feature 316 F.sub.n.sup.p and noisy proposal features 318 {circumflex over (F)}.sub.n.sup.p are sent to the SSPL 338. In an example, the SSPL 338 has two fully-connected layers with parameters θ.sup.self-loc and a sigmoid layer to compute location predictions Ln and noisy location predictions .sub.n.sup.p={{circumflex over (L)}.sub.nk.sup.p}.sub.k=1.sup.K. In an example, the numbers of the outputs of the two fully-connected layers are 1024 and 4 respectively.

(40) At process 380, a proposal location loss 340 L.sup.self-loc is computed, for example, as follows:

(41) $L^{\mathrm{self}‐\mathrm{loc}}\left(F_n^p,{\widehat{ℱ}}_n^p,P_n;{\theta }^{\mathrm{self}‐\mathrm{loc}}\right)=L^{\mathrm{self}‐\mathrm{loc}}\left(L_n^p,{\widehat{ℒ}}_n^p,P_n\right)=\frac{1}{K+1}\left({.Math.L_n^p-{\tilde{P}}_n.Math.}_2^2+\underset{k}{.Math.}{.Math.{\hat{L}}_{\mathrm{nk}}^p-{\tilde{P}}_n.Math.}_2^2\right).$

(42) In the example, l2 distance is to compute the proposal location loss 340 L.sup.self-loc, where

(43) {tilde over (P)}.sub.n=(x.sub.n/W, y.sub.n/H, w.sub.n/W, h.sub.n/H) is a normalized version of P.sub.n, and W and H denote the width and height of image I respectively.

(44) By optimizing this proposal location loss 340 L.sup.self-loc regarding proposal feature 316 F.sub.n.sup.p, noisy proposal features 318 {circumflex over (F)}.sub.n.sup.p, parameters θ.sup.seif-loc (i.e., θ.sup.b, θ.sup.r-cnn, θ.sup.self-loc), context-aware proposal features are learned, because predicting proposal location in an image requires proposal features understanding some global information of the image. In some embodiments, the relative patch location prediction task is not used directly, because when images are large and there are usually multiple objects in the same image for object detection, which makes relative patch location prediction hard to be solved.

(45) In some embodiments, the self-supervised proposal learning module uses a contrastive loss 342 L.sup.self-cont to learn noise-robust proposal features. For example, at step 382, the SSPL 338 may use a fully-connected layer with parameters θ.sup.self-cont and an l2 normalization layer to project proposal feature 316 F.sub.n.sup.p and noisy proposal features 318 {circumflex over (F)}.sub.n.sup.p to embedded features F.sub.n.sup.embed and {circumflex over (F)}.sub.n.sup.embed (e.g., .sub.n.sup.embed={{circumflex over (F)}.sub.nk.sup.embed}.sub.k=1.sup.K). In an example, the numbers of the outputs of the fully-connected layer is 128. An example of the contrastive loss 342 L.sup.self-cont may be written as follows:

(46) $L^{\mathrm{self}‐\mathrm{cont}}\left({\left\{F_n^p\right\}}_{n=1}^N,{\widehat{ℱ}}_n^p;{\theta }^{\mathrm{self}‐\mathrm{cont}}\right)=L^{\mathrm{self}‐\mathrm{cont}}\left({\left\{F_n^{\mathrm{embed}}\right\}}_{n=1}^N,{\widehat{ℱ}}_n^{\mathrm{embed}}\right)=-\frac{1}{K}\underset{k}{.Math.}\log \frac{\exp \left({\left({\hat{F}}_{\mathrm{nk}}^{\mathrm{embed}}\right)}^T{F}_n^{\mathrm{embed}}/\tau \right)}{{.Math.}_{n^{\prime }}\exp \left({\left({\hat{F}}_{\mathrm{nk}}^{\mathrm{embed}}\right)}^T{F}_{n^{\prime }}^{\mathrm{embed}}/\tau \right)}.$

(47) Here τ is a temperature hyper-parameter. By optimizing the contrastive loss 342 L.sup.self-cont with regard to proposal feature 316 F.sub.n.sup.p, noisy proposal features 318 {circumflex over (F)}.sub.n.sup.p, parameters θ.sup.self-cont (i.e., θ.sup.b, θ.sup.r-cnn, θ.sup.self-cont), noisy proposal features are enforced to be closer to their original proposal features than to other proposal features, which learns noise-robust proposal features and thus learns noise-robust object detectors.

(48) In some embodiments, an overall self-supervised proposal learning loss L.sup.self 344 may be generated using the proposal location loss 340 and the contrastive loss 342, e.g., as follows:

(49) $L^{\mathrm{self}}\left({\left\{F_n^p\right\}}_{n=1}^N,{\widehat{ℱ}}_n^p;{\theta }^{\mathrm{self}}\right)={\lambda }^{\mathrm{self}‐\mathrm{loc}}{L}^{\mathrm{self}‐\mathrm{loc}}\left(F_n^p,{\widehat{ℱ}}_n^p,P_n;{\theta }^{\mathrm{self}‐\mathrm{loc}}\right)+{\lambda }^{\mathrm{self}‐\mathrm{loc}}{L}^{\mathrm{self}‐\mathrm{loc}}\left({\left\{F_n^p\right\}}_{n=1}^N,{\widehat{ℱ}}_n^p;{\theta }^{\mathrm{self}‐\mathrm{cont}}\right).$

(50) Here λ.sup.self-loc and λ.sup.self-cont are loss weights, and θ.sup.self includes θ.sup.self-loc and θ.sup.self-cont. The overall self-supervised proposal learning loss L.sup.self 344 is optimized with regard to θ.sup.b, θ.sup.r-cnn, and θ.sup.self to learn proposal features.

(51) Consistency-Based Proposal Learning

(52) In some embodiments, consistency losses 346 L.sup.cons are applied to ensure consistency between noisy proposal predictions and their original proposal predictions. Different consistency losses may be applied to different predictions, including e.g., bounding box classification predictions and bounding box regression predictions. As shown in the example of FIG. 3, the consistency losses 346 L.sup.cons is generated based on original predictions 320 (e.g., obtained through process 398) and noisy predictions 322 (obtained through process 396).

(53) For a classification consistency loss 348 for bounding box classification predictions 322 C.sub.n.sup.p and the corresponding noisy bounding box classification predictions 334, the classification consistency loss L.sup.cons-cls may be computed, for example, as follows:

(54) $L^{\mathrm{cons}‐\mathrm{cls}}\left(C_n^p,{\widehat{𝒞}}_n^p\right)=\frac{1}{K}\underset{k}{.Math.}\mathrm{KL}\left(C_n^p.Math.{\hat{C}}_{\mathrm{nk}}^p\right).$

(55) In the particular example, KL divergence is used as the loss to enforce class predictions from noisy proposals and their original proposals to be consistent.

(56) Unlike image classification only having classification results, object detection also predicts object locations. To further ensure proposal prediction consistency, consistency loss is also computed to enforce object location predictions from noisy proposals and their original proposals to be consistent. For a regression consistency loss 350 L.sup.cons-reg for bounding box regression predictions 324 R.sub.n.sup.p and the corresponding noisy bounding box regression predictions 336, the regression consistency loss L.sup.cons-reg may be computed, for example, as follows:

(57) $L^{\mathrm{cons}‐\mathrm{reg}}\left(R_n^p,{\widehat{ℛ}}_n^p\right)=\underset{k}{\min }\left({\mathrm{smooth}}_{{\ell }_1}\left(R_n^p-{\widehat{ℛ}}_{\mathrm{nk}}^p\right)\right).$

(58) In the particular example, smoothed 11 loss is used, and the easiest noisy proposal feature is selected to compute the regression consistency loss L.sup.cons-reg for training stability.

(59) In some embodiments, the overall consistency based proposal learning loss 346 L.sup.cons is generated using the one or more consistency losses, for example, as follows:

(60) $L^{\mathrm{cons}}\left(C_n^p,R_n^p,{\widehat{𝒞}}_n^p,{\widehat{ℛ}}_n^p\right)={\lambda }^{\mathrm{cons}‐\mathrm{cls}}{L}^{\mathrm{cons}‐\mathrm{cls}}\left(C_n^p,{\widehat{𝒞}}_n^p\right)+{\lambda }^{\mathrm{cons}‐\mathrm{reg}}{L}^{\mathrm{cons}‐\mathrm{reg}}\left(R_n^p,{\widehat{ℛ}}_n^p\right).$

(61) Here λ.sup.cons-cls and λ.sup.cons-reg are loss weights. The overall consistency based proposal learning loss 346 L.sup.cons is optimized with regard to Ĉ.sub.n.sup.p, {circumflex over (R)}.sub.n.sup.p (not w.r.t. C.sub.n.sup.p, R.sub.n.sup.p), which is therefore optimized with regard to θ.sup.b, θ.sup.r-cnn, θ.sup.cls, and θ.sup.reg. As such, more noisy-robust proposal features and predictions are learned.

(62) In some embodiments, the neural network 300 including its object detectors are trained by applying the fully-supervised loss on labeled dataset D.sup.l, and apply the self-supervised proposal learning loss and the consistency-based proposal loss to unlabeled dataset D.sup.u. The combined loss L is optimized during the back-propagation process of the training, where the combined loss L may be determined, e.g., as follows:

(63) $L\left(I,𝒢;{\theta }^b,{\theta }^{\mathrm{rpn}},{\theta }^{r‐\mathrm{cnn}},{\theta }^{\mathrm{cls}},{\theta }^{\mathrm{reg}},{\theta }^{\mathrm{self}}\right)=\frac{1}{.Math.{𝒟}^1.Math.}\underset{\left(I,𝒢\right)\in {𝒟}^1}{.Math.}{L}^{\sup }\left(I,𝒢;{\theta }^b,{\theta }^{\mathrm{rpn}},{\theta }^{r‐\mathrm{cnn}},{\theta }^{\mathrm{cls}},{\theta }^{\mathrm{reg}}\right)+\frac{1}{.Math.{𝒟}^u.Math.}\underset{I\in {𝒟}^u}{.Math.}\frac{1}{N}\underset{n}{.Math.}{L}^{\mathrm{self}}\left({\left\{F_n^p\right\}}_{n=1}^N,{\widehat{ℱ}}_n^p,P_n;{\theta }^{\mathrm{self}}\right)+\frac{1}{.Math.{𝒟}^u.Math.}\underset{I\in {𝒟}^u}{.Math.}\frac{1}{N}\underset{n}{.Math.}{L}^{\mathrm{cons}}\left(C_n^p,R_n^p,{\widehat{𝒞}}_n^p,{\widehat{ℛ}}_n^p\right).$

(64) In some embodiments, the neural network 300 including its object detectors are trained by applying the fully-supervised loss on labeled dataset D.sup.l, and apply the self-supervised proposal learning loss and the consistency-based proposal loss to both labeled dataset D.sup.l and unlabeled dataset D.sup.u. The combined loss L is optimized during the back-propagation process of the training, where the combined loss L may be determined, e.g., as follows:

(65) $L\left(I,𝒢;{\theta }^b,{\theta }^{\mathrm{rpn}},{\theta }^{r‐\mathrm{cnn}},{\theta }^{\mathrm{cls}},{\theta }^{\mathrm{reg}},{\theta }^{\mathrm{self}}\right)=\frac{1}{.Math.{𝒟}^1.Math.}\underset{\left(I,𝒢\right)\in {𝒟}^1}{.Math.}{L}^{\sup }\left(I,𝒢;{\theta }^b,{\theta }^{\mathrm{rpn}},{\theta }^{r‐\mathrm{cnn}},{\theta }^{\mathrm{cls}},{\theta }^{\mathrm{reg}}\right)+\frac{1}{.Math.{𝒟}^1.Math.+.Math.{𝒟}^u.Math.}\underset{I\in {𝒟}^u,{𝒟}^u}{.Math.}\frac{1}{N}\underset{n}{.Math.}{L}^{\mathrm{self}}\left({\left\{F_n^p\right\}}_{n=1}^N,{\widehat{ℱ}}_n^p,P_n;{\theta }^{\mathrm{self}}\right)+\frac{1}{.Math.{𝒟}^1.Math.+.Math.{𝒟}^u.Math.}\underset{I\in {𝒟}^u,{𝒟}^u}{.Math.}\frac{1}{N}\underset{n}{.Math.}{L}^{\mathrm{cons}}\left(C_n^p,R_n^p,{\widehat{𝒞}}_n^p,{\widehat{ℛ}}_n^p\right).$

(66) Referring to the example of FIG. 4, illustrated is a trained neural network model 400 (e.g., trained neural network 300) for interference. As shown in FIG. 4, during inference, the proposal-based object detectors parts are used without using components like SSPL 338 or consistency computation, and as such, the proposal learning approach does not bring any extra inference computations.

(67) Some examples of computing devices, such as computing device 100 may include non-transitory, tangible, machine readable media that include executable code that when run by one or more processors (e.g., processor 110) may cause the one or more processors to perform the processes of method 200. Some common forms of machine readable media that may include the processes of method 200 are, for example, floppy disk, flexible disk, hard disk, magnetic tape, any other magnetic medium, CD-ROM, any other optical medium, punch cards, paper tape, any other physical medium with patterns of holes, RAM, PROM, EPROM, FLASH-EPROM, any other memory chip or cartridge, and/or any other medium from which a processor or computer is adapted to read.

(68) This description and the accompanying drawings that illustrate inventive aspects, embodiments, implementations, or applications should not be taken as limiting. Various mechanical, compositional, structural, electrical, and operational changes may be made without departing from the spirit and scope of this description and the claims. In some instances, well-known circuits, structures, or techniques have not been shown or described in detail in order not to obscure the embodiments of this disclosure Like numbers in two or more figures represent the same or similar elements.

(69) In this description, specific details are set forth describing some embodiments consistent with the present disclosure. Numerous specific details are set forth in order to provide a thorough understanding of the embodiments. It will be apparent, however, to one skilled in the art that some embodiments may be practiced without some or all of these specific details. The specific embodiments disclosed herein are meant to be illustrative but not limiting. One skilled in the art may realize other elements that, although not specifically described here, are within the scope and the spirit of this disclosure. In addition, to avoid unnecessary repetition, one or more features shown and described in association with one embodiment may be incorporated into other embodiments unless specifically described otherwise or if the one or more features would make an embodiment non-functional.

(70) Although illustrative embodiments have been shown and described, a wide range of modification, change and substitution is contemplated in the foregoing disclosure and in some instances, some features of the embodiments may be employed without a corresponding use of other features. One of ordinary skill in the art would recognize many variations, alternatives, and modifications. Thus, the scope of the invention should be limited only by the following claims, and it is appropriate that the claims be construed broadly and in a manner consistent with the scope of the embodiments disclosed herein.