System and method for generating a probability distribution of a location of an object
An object detection system for generating a probability distribution of the location of an object includes one or more processors and a memory in communication with the one or more processors. The memory includes an image acquisition module, a bounding box generator module, and a probability distribution generator module. The image acquisition module causes the one or more processors to obtain a two-dimension image displaying an object. The bounding box generator module causes the one or more processors to generate, using the two-dimensional image as an input, a bounding box of the object displayed in the two-dimensional image having a plurality of pixels. The probability distribution generator module causes the one or more processors to generate a probability distribution of a bounding box location for the object using a neural process using a pair of context points, a latent space, and a centered pixel location.
1. An object detection system comprising:
one or more processors; and
a memory in communication with the one or more processors, the memory having:
an image acquisition module, the image acquisition module having instructions, that, when executed by the one or more processors, cause the one or more processors to obtain a two-dimension image displaying an object,
a bounding box generator module, the bounding box generator module having instructions, that, when executed by the one or more processors, cause the one or more processors to generate, using the two-dimensional image as an input, a bounding box of the object displayed in the two-dimensional image, the bounding box having a plurality of pixels, and
a probability distribution generator module, the probability distribution generator module having instructions, that, when executed by the one or more processors, cause the one or more processors to generate for each pixel of the plurality of pixels forming the bounding box, a pixel location (x c ) and a bounding box value (y c ) of the object based on the pixel location (x c ), generate a probability distribution of a bounding box location (Y T ) for the object, wherein the probability distribution of the bounding box location (Y T ) is generated using a neural process using a pair of context points (y C , x C ), a latent space (z), and a centered pixel location (X T ), the centered pixel location (X T ) being a location of a center pixel of the bounding box.
2. The object detection system of claim 1 , wherein the pair of context points (y C , x C ) comprise a pixel location (x c ) and a probability (y C ) that a pixel at the pixel location (x c ) form part of the object.
3. The object detection system of claim 1 , wherein the probability distribution generator module further includes instructions, that, when executed by the one or more processors, cause the one or more processors to sample the latent space (z) from a known distribution.
4. The object detection system of claim 3 , wherein the known distribution is pre-selected as a parameter set used during training time.
5. The object detection system of claim 1 , wherein the bounding box is defined by a four-dimensional vector.
6. The object detection system of claim 1 , wherein the probability distribution of the bounding box location (Y T ) is a multinomial Gaussian distribution.
7. The object detection system of claim 6 , wherein the multinomial Gaussian distribution is expressed as:
Y T =N (μ,Σ),
where μ is a 4-dimensional position of the bounding box position and Σ is a 4×4 matrix describing the probability distribution around μ.
8. An object detection method comprising the steps of:
obtaining a two-dimension image displaying an object;
generating, using the two-dimensional image as an input, a bounding box of the object displayed in the two-dimensional image, the bounding box having a plurality of pixels;
generating for each pixel of the plurality of pixels forming the bounding box, a pixel location (x c ) and a bounding box value (y c ) of the object based on the pixel location (x c );
generating a probability distribution of a bounding box location (Y T ) for the object; and
wherein the probability distribution of the bounding box location (Y T ) is generated using a neural process using a pair of context points (y C , x C ), a latent space (z), and a centered pixel location (X T ), the centered pixel location (X T ) being a location of a center pixel of the bounding box.
9. The object detection method of claim 8 , wherein the pair of context points (y C , x C ) comprise a pixel location (x c ) and a probability (y C ) that a pixel at the pixel location (x c ) form part of the object.
10. The object detection method of claim 8 , further comprising the step of sampling the latent space (z) from a known distribution.
11. The object detection method of claim 10 , wherein the known distribution is pre-selected as a parameter set used during training time.
12. The object detection method of claim 8 , wherein the bounding box is defined by a four-dimensional vector.
13. The object detection method of claim 8 , wherein the probability distribution of the bounding box location (Y T ) is a multinomial Gaussian distribution.
14. The object detection method of claim 13 , wherein the multinomial Gaussian distribution is expressed as:
Y T =N (μ,Σ),
where μ is a 4-dimensional position of the bounding box position and Σ is a 4×4 matrix describing the probability distribution around μ.
15. A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to:
obtain a two-dimension image displaying an object;
generate, using the two-dimensional image as an input, a bounding box of the object displayed in the two-dimensional image, the bounding box having a plurality of pixels;
generate for each pixel of the plurality of pixels forming the bounding box, a pixel location (x c ) and a bounding box value (y c ) of the object based on the pixel location (x c );
generate a probability distribution of a bounding box location (Y T ) for the object; and
wherein the probability distribution of the bounding box location (Y T ) is generated using a neural process using a pair of context points (y C , x C ), a latent space (z), and a centered pixel location (X T ), the centered pixel location (X T ) being a location of a center pixel of the bounding box.
16. The non-transitory computer-readable medium of claim 15 , wherein the pair of context points (y C , x C ) comprise a pixel location (x c ) and a probability (y C ) that a pixel at the pixel location (x c ) form part of the object.
17. The non-transitory computer-readable medium of claim 15 , further comprising instructions, that, when executed by the one or more processors, cause the one or more processors to sample the latent space (z) from a known distribution.
18. The non-transitory computer-readable medium of claim 17 , wherein the known distribution is pre-selected as a parameter set used during training time.
19. The non-transitory computer-readable medium of claim 15 , wherein the probability distribution of the bounding box location (Y T ) is a multinomial Gaussian distribution.
20. The non-transitory computer-readable medium of claim 19 , wherein the multinomial Gaussian distribution is expressed as:
Y T =N (μ,Σ),
where μ is a 4-dimensional position of the bounding box position and Σ is a 4×4 matrix describing the probability distribution around μ.