System and method for video instance segmentation via multi-scale spatio-temporal split attention transformer
A system, method, computer readable storage medium for a computer vision system includes at least one video camera, and video processor circuitry. The method includes inputting a stream of video data and generating a sequence of image frames, segmenting and tracking, by the video analysis apparatus, object instances in the stream of video data, including receiving the sequence of image frames, analyzing the sequence of image frames using a video instance segmentation transformer to obtain a video instance mask sequence from the sequence of image frames, the transformer having a backbone network, a transformer encoder-decoder, and an instance matching and segmentation block, The encoder contains a multi-scale spatio-temporal split attention module to capture spatio-temporal feature relationships at multiple scales across multiple frames. The decoder contains a temporal attention block for enhancing a temporal consistency of transformer queries. The method includes displaying the video instance mask sequence.
1 . A vehicle computer vision system, comprising:
at least one video camera mounted on a vehicle to capture a stream of video data;
video processor circuitry configured to input the stream of video data and generate a sequence of image frames;
segment and track object instances in the stream of video data, including
receiving the sequence of image frames;
obtaining a video instance mask sequence from the sequence of image frames, via a video instance segmentation transformer having
a backbone network,
a transformer encoder-decoder, and
an instance matching and segmentation block,
wherein the encoder contains a multi-scale spatio-temporal split (MS-STS) attention module to capture spatio-temporal feature relationships at multiple scales across multiple frames of the sequence of image frames, the MT-STS attention module producing multi-scale spatio-temporal enriched features;
wherein the decoder contains a temporal attention block for enhancing a temporal consistency of transformer queries;
wherein the encoder further contains a deformable attention block that outputs baseline features, the baseline features being fused with the multi-scale spatio-temporal enriched features through a convolution operation to obtain encoder output features; and
display a sequence of mask instances.
2 . The vehicle computer vision system of claim 1 , wherein
the MS-STS attention module of the video processor circuitry employs an intra-scale temporal attention block and an inter-scale temporal attention block;
the intra-scale temporal attention block of the video processor circuitry operates on each spatial scale across the multiple frames and produces temporally relevant intermediate features; and
the inter-scale temporal attention block of the video processor circuitry takes the temporally relevant intermediate features output by the intra-scale temporal attention block and learns similarities between the spatial-temporal feature relationships across two neighboring spatial scales.
3 . The vehicle computer vision system of claim 1 , wherein the video processor circuitry is further configured to
during training, use an adversarial loss to improve the spatio-temporal feature relationships which enhances foreground-background separability.
4 . The vehicle computer vision system of claim 1 , wherein
an output of the encoder and the temporally consistent transformer queries from the decoder are used within the instance matching and segmentation block of the video processor circuitry for prediction of the sequence of mask instances.
5 . The vehicle computer vision system of claim 1 , wherein
the temporal attention block of the decoder is connected between a self-attention block and a cross-attention block of the decoder.
6 . The vehicle computer vision system of claim 1 , wherein
the received sequence of image frames includes a sequence of image frames of a road sign of different scales, in motion; and
the multi-scale spatio-temporal split attention module captures the spatio-temporal feature relationships at multiple scales across multiple frames that includes the road sign as the vehicle approaches and passes the road sign.
7 . The vehicle computer vision system of claim 1 , wherein
the received sequence of video image frames includes a sequence of image frames of a road sign of different scales, in motion, and at least one pedestrian also at different scales moving in a direction of a crosswalk; and
the multi-scale spatio-temporal split attention module captures the spatio-temporal feature relationships at multiple scales across multiple frames that includes the road sign as the vehicle approaches the road sign together with the at least one pedestrian moving toward the crosswalk.
8 . The vehicle computer vision system of claim 1 , wherein
the received sequence of video image frames includes a dynamic scene in which people are moving in random directions in a vicinity of a road which the vehicle is traveling, crossing paths with each other;
the multi-scale spatio-temporal split attention module captures the spatio-temporal feature relationships at multiple scales across multiple frames that includes the moving people as the vehicle approaches the moving people; and
the vehicle computer vision system tracks the moving people.
9 . The vehicle computer vision system of claim 1 , wherein
the received sequence of video image frames includes a scene containing a location road sign that displays information of a location;
the multi-scale spatio-temporal split attention module
captures the spatio-temporal feature relationships at multiple scales across multiple frames that includes the information of the location road sign, under conditions of changing scale of the location road sign, subject to various angles of view and partial view of the location road sign, and
recognizes the road sign including the information of the road sign, while the vehicle is moving, under conditions of changing scale of the location road sign, while the vehicle moves relative to the location road sign, and subject to various angles of view and partial view of the location road sign.
10 . The vehicle computer vision system of claim 1 , wherein
the received sequence of video image frames includes a partial scene of a road where the vehicle is traveling that includes a plurality of moving pedestrians entering and leaving the partial scene, at least one cyclist moving at a higher rate of speed relative to the plurality of pedestrians, into the partial scene, and a partial view of a building;
the multi-scale spatio-temporal split attention module captures the spatio-temporal feature relationships at various scales across frames in the video sequence that includes the plurality of pedestrians and the at least one cyclist as the vehicle approaches and passes a section of the road where the pedestrians and the at least one cyclist are approaching to cross; and
the vehicle computer vision system tracks the moving plurality of pedestrians and at least one cyclist.
11 . A non-transitory computer-readable storage medium storing computer program instructions, which when executed in a vehicle computer vision system, the vehicle computer vision system comprising at least one video camera mounted on a vehicle, and video processor circuitry, perform a method comprising:
inputting a stream of video data and generating a sequence of image frames;
segmenting and tracking, by the video processor circuitry, object instances in the stream of video data, including
receiving the sequence of image frames;
analyzing the sequence of image frames using a video instance segmentation transformer to obtain a video instance mask sequence from the sequence of image frames, the transformer having
a backbone network,
a transformer encoder-decoder, and
an instance matching and segmentation block,
wherein the encoder contains a multi-scale spatio-temporal split (MS-STS) attention module to capture spatio-temporal feature relationships at multiple scales across multiple frames, the MT-STS attention module producing multi-scale spatio-temporal enriched features;
wherein the decoder contains a temporal attention block for enhancing a temporal consistency of transformer queries;
wherein the encoder further contains a deformable attention block that outputs baseline features, the baseline features being fused with the multi-scale spatio-temporal enriched features through a convolution operation to obtain encoder output features; and
displaying the video instance mask sequence.
12 . The computer readable storage medium of claim 11 , wherein
the MS-STS attention module employs an intra-scale temporal attention block and an inter-scale temporal attention block, further comprising:
operating, via the intra-scale temporal attention block, on each spatial scale across the multiple frames to produce temporally relevant intermediate features; and
learning, via the inter-scale temporal attention block, similarities between the spatial-temporal feature relationships across two neighboring spatial scales using the temporally relevant intermediate features output by the intra-scale temporal attention block.
13 . The computer readable storage medium of claim 11 , further comprising:
training the video processor circuitry, including improving the spatio-temporal feature relationships using an adversarial loss which enhances foreground-background separability.
14 . The computer readable storage medium of claim 11 , further comprising:
predicting the sequence of mask instances using the output of the encoder and the temporally consistent transformer queries from the decoder.
15 . The computer readable storage medium of claim 11 , wherein:
the video image frame sequence includes a sequence of image frames of a road sign of different scales, in motion; further comprising:
capturing, via the multi-scale spatio-temporal split attention module, the spatio-temporal feature relationships at multiple scales across multiple frames that includes the road sign as the vehicle approaches and passes the road sign.
16 . The computer readable storage medium of claim 11 , wherein
the video image frame sequence includes a sequence of image frames of a road sign of different scales, in motion, and at least one pedestrian also at different scales moving in a direction of a crosswalk; further comprising:
capturing, via the multi-scale spatio-temporal split attention module, the spatio-temporal feature relationships at multiple scales across multiple frames that includes the road sign as the vehicle approaches the road sign together with the at least one pedestrian moving toward the crosswalk.
17 . The computer readable storage medium of claim 11 , wherein
the video image frame sequence includes a dynamic scene in which people are moving in random directions in a vicinity of a road which the vehicle is traveling, crossing paths with each other; further comprising:
capturing, via the multi-scale spatio-temporal split attention module, the spatio-temporal feature relationships at multiple scales across multiple frames that includes the moving people as the vehicle approaches the moving people; and
tracking the moving people.
18 . The computer readable storage medium of claim 11 , wherein
the video image frame sequence includes a scene containing a location road sign that displays information of a location; further comprising:
capturing, via the multi-scale spatio-temporal split attention module, the spatio-temporal feature relationships at multiple scales across multiple frames that includes the information of the location road sign, under conditions of changing scale of the location road sign, subject to various angles of view and partial view of the location road sign; and
recognizing the road sign including the information of the road sign, while the vehicle is moving, under conditions of changing scale of the location road sign, while the vehicle moves relative to the location road sign, and subject to various angles of view and partial view of the location road sign.
19 . The computer readable storage medium of claim 11 , wherein
the video image frame sequence includes a partial scene of a road where the vehicle is traveling that includes a plurality of moving pedestrians entering and leaving the partial scene, at least one cyclist moving at a higher rate of speed relative to the plurality of pedestrians, into the partial scene, and a partial view of a building; further comprising:
capturing, via the multi-scale spatio-temporal split attention module, the spatio-temporal feature relationships at various scales across frames in the video sequence that includes the plurality of pedestrians and the at least one cyclist as the vehicle approaches and passes a section of the road where the pedestrians and the at least one cyclist are approaching to cross; and
tracking the moving plurality of pedestrians and at least one cyclist.
20 . A method for a vehicle computer vision system, the vehicle computer vision system comprising at least one video camera mounted on a vehicle, and video processor circuitry, the method comprising:
inputting a stream of video data and generating a sequence of image frames;
segmenting and tracking, by the video processor circuitry, object instances in the stream of video data, including
receiving the sequence of image frames;
analyzing the sequence of image frames using a video instance segmentation transformer to obtain a video instance mask sequence from the sequence of image frames, having
a backbone network,
a transformer encoder-decoder, and
an instance matching and segmentation block,
wherein the encoder contains a multi-scale spatio-temporal split (MS-STS) attention module to capture spatio-temporal feature relationships at multiple scales across multiple frames, the MT-STS attention module producing multi-scale spatio-temporal enriched features;
wherein the decoder contains a temporal attention block for enhancing a temporal consistency of transformer queries;
wherein the encoder further contains a deformable attention block that outputs baseline features, the baseline features being fused with the multi-scale spatio-temporal enriched features through a convolution operation to obtain encoder output features; and
displaying the video instance mask sequence.