Continual 3D Convolutional Neural Networks for Real-time Processing of Videos

31 May 2021  ·  Lukas Hedegaard, Alexandros Iosifidis ·

We introduce Continual 3D Convolutional Neural Networks (Co3D CNNs), a new computational formulation of spatio-temporal 3D CNNs, in which videos are processed frame-by-frame rather than by clip. In online tasks demanding frame-wise predictions, Co3D CNNs dispense with the computational redundancies of regular 3D CNNs, namely the repeated convolutions over frames, which appear in overlapping clips. We show that Continual 3D CNNs can reuse preexisting 3D-CNN weights to reduce the per-prediction floating point operations (FLOPs) in proportion to the temporal receptive field while retaining similar memory requirements and accuracy. This is validated with multiple models on Kinetics-400 and Charades with remarkable results: CoX3D models attain state-of-the-art complexity/accuracy trade-offs on Kinetics-400 with 12.1-15.3x reductions of FLOPs and 2.3-3.8% improvements in accuracy compared to regular X3D models while reducing peak memory consumption by up to 48%. Moreover, we investigate the transient response of Co3D CNNs at start-up and perform extensive benchmarks of on-hardware processing characteristics for publicly available 3D CNNs.

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Action Classification Charades Co Slow_64 MAP 25.2 # 42
FLOPs (G) x views 6.9x1 # 1
Action Classification Charades Co Slow_8 MAP 21.5 # 47
FLOPs (G) x views 6.9x1 # 1
Action Classification Charades Slow-8×8 MAP 24.1 # 44
FLOPs (G) x views 54.9x1 # 1
Action Classification Kinetics-400 Co X3D-S_64 Acc@1 67.33 # 180
FLOPs (G) x views 0.17x1 # 1
Parameters (M) 3.79 # 1
Action Classification Kinetics-400 Co X3D-S_13 Acc@1 60.18 # 189
FLOPs (G) x views 0.17x1 # 1
Parameters (M) 3.79 # 1
Action Classification Kinetics-400 Co X3D-M_64 Acc@1 71.03 # 173
FLOPs (G) x views 0.33x1 # 1
Parameters (M) 3.79 # 1
Action Classification Kinetics-400 Co X3D-M_16 Acc@1 62.80 # 188
FLOPs (G) x views 0.33x1 # 1
Parameters (M) 3.79 # 1
Action Classification Kinetics-400 Co X3D-L_64 Acc@1 71.61 # 171
FLOPs (G) x views 1.25x1 # 1
Parameters (M) 6.15 # 8
Action Classification Kinetics-400 Co X3D-L_16 Acc@1 63.03 # 187
FLOPs (G) x views 1.25x1 # 1
Parameters (M) 6.15 # 8
Action Classification Kinetics-400 Co Slow_64 Acc@1 73.05 # 162
FLOPs (G) x views 6.90x1 # 1
Parameters (M) 32.45 # 18
Action Classification Kinetics-400 Co Slow_8 Acc@1 65.90 # 183
FLOPs (G) x views 6.90x1 # 1
Parameters (M) 32.45 # 18
Action Classification Kinetics-400 Co I3D_64 Acc@1 56.86 # 193
FLOPs (G) x views 5.68x1 # 1
Parameters (M) 28.04 # 13
Action Classification Kinetics-400 Co I3D_8 Acc@1 59.58 # 190
FLOPs (G) x views 5.68x1 # 1
Parameters (M) 28.04 # 13
Action Classification Kinetics-400 RCU_8 Acc@1 53.40 # 195
FLOPs (G) x views 4.71x1 # 1
Parameters (M) 12.80 # 11
Action Classification Kinetics-400 X3D-XS Acc@1 59.37 # 192
FLOPs (G) x views 0.64x1 # 1
Parameters (M) 3.79 # 1
Action Classification Kinetics-400 X3D-S Acc@1 64.71 # 185
FLOPs (G) x views 2.06x1 # 1
Parameters (M) 3.79 # 1
Action Classification Kinetics-400 X3D-M Acc@1 67.24 # 181
FLOPs (G) x views 4.97x1 # 1
Parameters (M) 3.79 # 1
Action Classification Kinetics-400 X3D-L Acc@1 69.29 # 174
FLOPs (G) x views 19.17x1 # 1
Parameters (M) 6.15 # 8
Action Classification Kinetics-400 SlowFast-4×16-R50 Acc@1 67.06 # 182
FLOPs (G) x views 36.46x1 # 1
Parameters (M) 34.48 # 21
Action Classification Kinetics-400 SlowFast-8×8-R50 Acc@1 68.45 # 176
FLOPs (G) x views 66.25x1 # 1
Parameters (M) 66.25 # 22
Action Classification Kinetics-400 Slow-8x8-R50 Acc@1 67.42 # 179
FLOPs (G) x views 54.87x1 # 1
Parameters (M) 32.45 # 18
Action Classification Kinetics-400 R(2+1)D-18_16 Acc@1 59.52 # 191
FLOPs (G) x views 40.71x1 # 1
Parameters (M) 31.51 # 16
Action Classification Kinetics-400 R(2+1)D-18_8 Acc@1 53.52 # 194
FLOPs (G) x views 20.35x1 # 1
Parameters (M) 31.51 # 16
Action Classification Kinetics-400 I3D-R50 Acc@1 63.98 # 186
FLOPs (G) x views 28.61x1 # 1
Parameters (M) 28.04 # 13

Methods


No methods listed for this paper. Add relevant methods here