Systems and methods for improving object tracking in compressed feature data in coding of multi-dimensional data
A method of compressing feature data includes: receiving feature data; performing spatial down sampling on the received feature data by applying a pixel unshuffle operation; and performing channel reduction on the spatially down sampled feature data by applying a non-linear two dimensional convolution with an activation.
1. A method of compressing feature data, the method comprising:
receiving W×H×C feature data, wherein W and H are spatial dimensions and C is a number of channels and each channel corresponds to a down sampled version of an input image;
performing spatial down sampling on the received W×H×C feature data by applying a pixel unshuffle operation using a downscale factor to generate W/r×H/r×C spatially down sampled feature data, wherein r is the downscale factor;
reducing the number of channels from C to N by applying a non-linear two dimensional convolution with an activation and using a residual connection;
applying a spatial batch normalization layer to W/r×H/r×N feature data;
quantizing the W/r×H/r×N feature data; and
entropy coding the quantized W/r×H/r×N feature data to generate a bitstream.
2. The method of claim 1 , wherein r is equal to 3, C is equal to 256, and N is equal to 16.
3. The method of claim 1 , wherein the received W×H×C feature data includes residual data generated using inter prediction.