| Output: Weight matrices W_and A. | Input: Samples (x1, y1),.,(xn, yn) generated according to Equation (4) | |
| 1: {Finding span of vec(zizT)’s} | ||
| 2: Estimate empirical moments E[xxT],E[yxT],E[yx@3] and E[y y (x x)]. | ||
| 3: Compute the vector f(u) = 2E[(uTy).(E[(uTy)xT]E[xxT]-1x)·(xx)]-E[(uTy)²(x凶x)] | ||
| where each entry is expressed as a degree-2 polynomial over u. | ||
| 4: | : Construct matrix T ∈ Rd²×(k2+k) such that,Tvec*(uuT) = f(u). | |
| 5: | Compute the k-least right singular vectors of T,denoted by vec*(U1), vec*(U2),.: , vec*(Uk). Let S be a k2 + k by k matrix, where the i-th column of S is vector vec*(Ui). {Here span S is | |
| equal to span of {vec*(ziz𝑖)}.} 6:{Finding zi's from span} | ||
| 7: Let X = mat*(SSi)， Y = mat*(SS2)， where S1 and S2 are two independently sampled k- | ||
| 8: | dimensional standard Gaussian vectors. :Let z1,.,zk be eigenvectors of XY-1. | |
| 9: | For each zi,ifE[zTy] <O let zi ← -zi. {Here zi's are normalized rows of A-1.} | |
| 10: | {Reduce to 1-Layer Problem} | |
| 11: | For each zi,let Ui be the output of Algorithm 1 with input (x1,zi y1),., (xn, zi yn). | |
| 13:return V,Z-1. | 12: Let Z be the matrix whose rows are zT's, V be the matrix whose rows are vi 's. |
| Output: Weight matrices W_and A. | Input: Samples (x1, y1),.,(xn, yn) generated according to Equation (4) | |
| 1: {Finding span of vec(zizT)’s} | ||
| 2: Estimate empirical moments E[xxT],E[yxT],E[yx@3] and E[y y (x x)]. | ||
| 3: Compute the vector f(u) = 2E[(uTy).(E[(uTy)xT]E[xxT]-1x)·(xx)]-E[(uTy)²(x凶x)] | ||
| where each entry is expressed as a degree-2 polynomial over u. | ||
| 4: | : Construct matrix T ∈ Rd²×(k2+k) such that,Tvec*(uuT) = f(u). | |
| 5: | Compute the k-least right singular vectors of T,denoted by vec*(U1), vec*(U2),.: , vec*(Uk). Let S be a k2 + k by k matrix, where the i-th column of S is vector vec*(Ui). {Here span S is | |
| equal to span of {vec*(ziz𝑖)}.} 6:{Finding zi's from span} | ||
| 7: Let X = mat*(SSi)， Y = mat*(SS2)， where S1 and S2 are two independently sampled k- | ||
| 8: | dimensional standard Gaussian vectors. :Let z1,.,zk be eigenvectors of XY-1. | |
| 9: | For each zi,ifE[zTy] <O let zi ← -zi. {Here zi's are normalized rows of A-1.} | |
| 10: | {Reduce to 1-Layer Problem} | |
| 11: | For each zi,let Ui be the output of Algorithm 1 with input (x1,zi y1),., (xn, zi yn). | |
| 13:return V,Z-1. | 12: Let Z be the matrix whose rows are zT's, V be the matrix whose rows are vi 's. |
| Output: Weight matrices W_and A. | Input: Samples (x1, y1),.,(xn, yn) generated according to Equation (4) | |
| 1: {Finding span of vec(zizT)’s} | ||
| 2: Estimate empirical moments E[xxT],E[yxT],E[yx@3] and E[y y (x x)]. | ||
| 3: Compute the vector f(u) = 2E[(uTy).(E[(uTy)xT]E[xxT]-1x)·(xx)]-E[(uTy)²(x凶x)] | ||
| where each entry is expressed as a degree-2 polynomial over u. | ||
| 4: | : Construct matrix T ∈ Rd²×(k2+k) such that,Tvec*(uuT) = f(u). | |
| 5: | Compute the k-least right singular vectors of T,denoted by vec*(U1), vec*(U2),.: , vec*(Uk). Let S be a k2 + k by k matrix, where the i-th column of S is vector vec*(Ui). {Here span S is | |
| equal to span of {vec*(ziz𝑖)}.} 6:{Finding zi's from span} | ||
| 7: Let X = mat*(SSi)， Y = mat*(SS2)， where S1 and S2 are two independently sampled k- | ||
| 8: | dimensional standard Gaussian vectors. :Let z1,.,zk be eigenvectors of XY-1. | |
| 9: | For each zi,ifE[zTy] <O let zi ← -zi. {Here zi's are normalized rows of A-1.} | |
| 10: | {Reduce to 1-Layer Problem} | |
| 11: | For each zi,let Ui be the output of Algorithm 1 with input (x1,zi y1),., (xn, zi yn). | |
| 13:return V,Z-1. | 12: Let Z be the matrix whose rows are zT's, V be the matrix whose rows are vi 's. |
| Output: Weight matrices W_and A. | Input: Samples (x1, y1),.,(xn, yn) generated according to Equation (4) | |
| 1: {Finding span of vec(zizT)’s} | ||
| 2: Estimate empirical moments E[xxT],E[yxT],E[yx@3] and E[y y (x x)]. | ||
| 3: Compute the vector f(u) = 2E[(uTy).(E[(uTy)xT]E[xxT]-1x)·(xx)]-E[(uTy)²(x凶x)] | ||
| where each entry is expressed as a degree-2 polynomial over u. | ||
| 4: | : Construct matrix T ∈ Rd²×(k2+k) such that,Tvec*(uuT) = f(u). | |
| 5: | Compute the k-least right singular vectors of T,denoted by vec*(U1), vec*(U2),.: , vec*(Uk). Let S be a k2 + k by k matrix, where the i-th column of S is vector vec*(Ui). {Here span S is | |
| equal to span of {vec*(ziz𝑖)}.} 6:{Finding zi's from span} | ||
| 7: Let X = mat*(SSi)， Y = mat*(SS2)， where S1 and S2 are two independently sampled k- | ||
| 8: | dimensional standard Gaussian vectors. :Let z1,.,zk be eigenvectors of XY-1. | |
| 9: | For each zi,ifE[zTy] <O let zi ← -zi. {Here zi's are normalized rows of A-1.} | |
| 10: | {Reduce to 1-Layer Problem} | |
| 11: | For each zi,let Ui be the output of Algorithm 1 with input (x1,zi y1),., (xn, zi yn). | |
| 13:return V,Z-1. | 12: Let Z be the matrix whose rows are zT's, V be the matrix whose rows are vi 's. |
| Input: Samples (x1, y1),., (xn,yn) generated according to Equation (3). | |||
| Output: Estimate of weight vector w. n | |||
| 1:Estimate U = | 1 ∑i=1yixi. n | ||
| 2:Estimate C= 3: return 2C-1v. | Ci=1xixi T |
| Output: Weight matrices W_and A. | Input: Samples (x1, y1),.,(xn, yn) generated according to Equation (4) | |
| 1: {Finding span of vec(zizT)’s} | ||
| 2: Estimate empirical moments E[xxT],E[yxT],E[yx@3] and E[y y (x x)]. | ||
| 3: Compute the vector f(u) = 2E[(uTy).(E[(uTy)xT]E[xxT]-1x)·(xx)]-E[(uTy)²(x凶x)] | ||
| where each entry is expressed as a degree-2 polynomial over u. | ||
| 4: | : Construct matrix T ∈ Rd²×(k2+k) such that,Tvec*(uuT) = f(u). | |
| 5: | Compute the k-least right singular vectors of T,denoted by vec*(U1), vec*(U2),.: , vec*(Uk). Let S be a k2 + k by k matrix, where the i-th column of S is vector vec*(Ui). {Here span S is | |
| equal to span of {vec*(ziz𝑖)}.} 6:{Finding zi's from span} | ||
| 7: Let X = mat*(SSi)， Y = mat*(SS2)， where S1 and S2 are two independently sampled k- | ||
| 8: | dimensional standard Gaussian vectors. :Let z1,.,zk be eigenvectors of XY-1. | |
| 9: | For each zi,ifE[zTy] <O let zi ← -zi. {Here zi's are normalized rows of A-1.} | |
| 10: | {Reduce to 1-Layer Problem} | |
| 11: | For each zi,let Ui be the output of Algorithm 1 with input (x1,zi y1),., (xn, zi yn). | |
| 13:return V,Z-1. | 12: Let Z be the matrix whose rows are zT's, V be the matrix whose rows are vi 's. |
| Method | #Attributes | Coverage (%) | Correctness (%) | |
| iid split | ||||
| TELBO | 4 | 82.08±0.56 | ||
| JMVAE | 85.15 ±0.26 | |||
| BiVCCA | 67.38 ±0.69 | |||
| TELBO | 3 | 91.14 ± 0.53 | 81.63 ± 0.38 | |
| JMVAE | 88.52 ± 0.37 | 82.00 ±0.37 | ||
| BiVCCA | 85.28 ±0.68 | 70.68 ± 0.87 | ||
| TELBO | 2 | 90.32 ± 0.57 | 82.03 ± 1.37 | |
| JMVAE | 87.89 ±0.69 | 81.02 ± 1.05 | ||
| BiVCCA | 85.09 ±0.76 | 72.33 ± 2.31 | ||
| TELBO | 1 | 90.94 ±0.19 | 83.67 ± 1.70 | |
| JMVAE | 88.70 ±0.35 | 81.58 ± 1.78 | ||
| BiVCCA | 85.53 ± 0.27 | 68.36 ± 2.21 | ||
| Compositional split | ||||
| TELBO | 4 | 75.61 ± 1.43 | ||
| JMVAE | 76.86 ± 1.30 | |||
| BiVCCA | 68.58 ± 1.02 | |||
| Approach | Concept-Latent (%) | Concept-NB (%) |
| TELBO | 35.66 ± 2.05 | 17.66 ± 1.70 |
| JMVAE | 54.66 ± 4.92 | 13.33 ± 2.05 |
| BiVCCA | 28.00 ± 4.54 | 18.00 ± 1.40 |
| Random | 0.28±0.00 | 0.28±0.00 |
| Most Frequent | 6.33 ± 1.88 | 6.33 ± 1.88 |
| Method | #Attributes | Coverage (%) | Correctness (%) |
| iid split | |||
| TELBO | 4 | 82.08±0.56 | |
| JMVAE | 85.15 ±0.26 | ||
| BiVCCA | 67.38 ±0.69 | ||
| TELBO | 3 | 91.14 ± 0.53 | 81.63 ± 0.38 |
| JMVAE | 88.52 ± 0.37 | 82.00 ±0.37 | |
| BiVCCA | 85.28 ±0.68 | 70.68 ± 0.87 | |
| TELBO | 2 | 90.32 ± 0.57 | 82.03 ± 1.37 |
| JMVAE | 87.89 ±0.69 | 81.02 ± 1.05 | |
| BiVCCA | 85.09 ±0.76 | 72.33 ± 2.31 | |
| TELBO | 1 | 90.94 ±0.19 | 83.67 ± 1.70 |
| JMVAE | 88.70 ±0.35 | 81.58 ± 1.78 | |
| BiVCCA | 85.53 ± 0.27 | 68.36 ± 2.21 | |
| Compositional split | |||
| TELBO | 4 | 75.61 ± 1.43 | |
| JMVAE | 76.86 ± 1.30 | ||
| BiVCCA | 68.58 ± 1.02 |
| Approach | Concept-Latent (%) | Concept-NB (%) |
| TELBO | 35.66 ± 2.05 | 17.66 ± 1.70 |
| JMVAE | 54.66 ± 4.92 | 13.33 ± 2.05 |
| BiVCCA | 28.00 ± 4.54 | 18.00 ± 1.40 |
| Random | 0.28±0.00 | 0.28±0.00 |
| Most Frequent | 6.33 ± 1.88 | 6.33 ± 1.88 |
| Method | #Attributes | Coverage (%) | Correctness (%) |
| iid split | |||
| TELBO | 4 | 82.08±0.56 | |
| JMVAE | 85.15 ±0.26 | ||
| BiVCCA | 67.38 ±0.69 | ||
| TELBO | 3 | 91.14 ± 0.53 | 81.63 ± 0.38 |
| JMVAE | 88.52 ± 0.37 | 82.00 ±0.37 | |
| BiVCCA | 85.28 ±0.68 | 70.68 ± 0.87 | |
| TELBO | 2 | 90.32 ± 0.57 | 82.03 ± 1.37 |
| JMVAE | 87.89 ±0.69 | 81.02 ± 1.05 | |
| BiVCCA | 85.09 ±0.76 | 72.33 ± 2.31 | |
| TELBO | 1 | 90.94 ±0.19 | 83.67 ± 1.70 |
| JMVAE | 88.70 ±0.35 | 81.58 ± 1.78 | |
| BiVCCA | 85.53 ± 0.27 | 68.36 ± 2.21 | |
| Compositional split | |||
| TELBO | 4 | 75.61 ± 1.43 | |
| JMVAE | 76.86 ± 1.30 | ||
| BiVCCA | 68.58 ± 1.02 |
| Method | #Attributes | Coverage (%) | Correctness (%) |
| iid split | |||
| TELBO | 4 | 82.08±0.56 | |
| JMVAE | 85.15 ±0.26 | ||
| BiVCCA | 67.38 ±0.69 | ||
| TELBO | 3 | 91.14 ± 0.53 | 81.63 ± 0.38 |
| JMVAE | 88.52 ± 0.37 | 82.00 ±0.37 | |
| BiVCCA | 85.28 ±0.68 | 70.68 ± 0.87 | |
| TELBO | 2 | 90.32 ± 0.57 | 82.03 ± 1.37 |
| JMVAE | 87.89 ±0.69 | 81.02 ± 1.05 | |
| BiVCCA | 85.09 ±0.76 | 72.33 ± 2.31 | |
| TELBO | 1 | 90.94 ±0.19 | 83.67 ± 1.70 |
| JMVAE | 88.70 ±0.35 | 81.58 ± 1.78 | |
| BiVCCA | 85.53 ± 0.27 | 68.36 ± 2.21 | |
| Compositional split | |||
| TELBO | 4 | 75.61 ± 1.43 | |
| JMVAE | 76.86 ± 1.30 | ||
| BiVCCA | 68.58 ± 1.02 |
| Approach | Concept-Latent (%) | Concept-NB (%) |
| TELBO | 35.66 ± 2.05 | 17.66 ± 1.70 |
| JMVAE | 54.66 ± 4.92 | 13.33 ± 2.05 |
| BiVCCA | 28.00 ± 4.54 | 18.00 ± 1.40 |
| Random | 0.28±0.00 | 0.28±0.00 |
| Most Frequent | 6.33 ± 1.88 | 6.33 ± 1.88 |
| Approach | Concept-Latent (%) | Concept-NB (%) |
| TELBO | 35.66 ± 2.05 | 17.66 ± 1.70 |
| JMVAE | 54.66 ± 4.92 | 13.33 ± 2.05 |
| BiVCCA | 28.00 ± 4.54 | 18.00 ± 1.40 |
| Random | 0.28±0.00 | 0.28±0.00 |
| Most Frequent | 6.33 ± 1.88 | 6.33 ± 1.88 |
| Method | #Attributes | Coverage (%) | Correctness (%) |
| iid split | |||
| TELBO | 4 | 82.08±0.56 | |
| JMVAE | 85.15 ±0.26 | ||
| BiVCCA | 67.38 ±0.69 | ||
| TELBO | 3 | 91.14 ± 0.53 | 81.63 ± 0.38 |
| JMVAE | 88.52 ± 0.37 | 82.00 ±0.37 | |
| BiVCCA | 85.28 ±0.68 | 70.68 ± 0.87 | |
| TELBO | 2 | 90.32 ± 0.57 | 82.03 ± 1.37 |
| JMVAE | 87.89 ±0.69 | 81.02 ± 1.05 | |
| BiVCCA | 85.09 ±0.76 | 72.33 ± 2.31 | |
| TELBO | 1 | 90.94 ±0.19 | 83.67 ± 1.70 |
| JMVAE | 88.70 ±0.35 | 81.58 ± 1.78 | |
| BiVCCA | 85.53 ± 0.27 | 68.36 ± 2.21 | |
| Compositional split | |||
| TELBO | 4 | 75.61 ± 1.43 | |
| JMVAE | 76.86 ± 1.30 | ||
| BiVCCA | 68.58 ± 1.02 |
| Approach | Concept-Latent (%) | Concept-NB (%) |
| TELBO | 35.66 ± 2.05 | 17.66 ± 1.70 |
| JMVAE | 54.66 ± 4.92 | 13.33 ± 2.05 |
| BiVCCA | 28.00 ± 4.54 | 18.00 ± 1.40 |
| Random | 0.28±0.00 | 0.28±0.00 |
| Most Frequent | 6.33 ± 1.88 | 6.33 ± 1.88 |
| Method | Backflip | Cartwheel | Dance | Run | Spinkick |
| BC | 3.01 | 2.88 | 2.93 | 2.63 | 2.88 |
| Merel et al., 2017 | 1.33 ± 0.03 | 1.47 ± 0.12 | 2.61 ± 0.30 | 0.52 ± 0.04 | 1.82 ± 0.35 |
| GAIL | 0.74±0.15 | 0.84± 0.05 | 1.31 ± 0.16 | 0.17±0.03 | 1.07 ± 0.03 |
| GAIL -noise | 0.42±0.02 | 0.92±0.07 | 0.96±0.08 | 0.21±0.05 | 0.95±0.14 |
| GAIL - noise z | 0.67±0.12 | 0.72± 0.04 | 1.14 ± 0.08 | 0.14±0.03 | 0.64±0.09 |
| GAIL - GP | 0.62±0.09 | 0.69 ±0.05 | 0.80± 0.32 | 0.12 ± 0.02 | 0.64± 0.04 |
| VAIL (ours) | 0.36±0.13 | 0.40±0.08 | 0.40±0.21 | 0.13±0.01 | 0.34± 0.05 |
| VAIL - GP (ours) | 0.46±0.17 | 0.31 ± 0.02 | 0.15±0.01 | 0.10±0.01 | 0.31 ± 0.02 |
| Peng et al., 2018 | 0.26 | 0.21 | 0.20 | 0.14 | 0.19 |
| Method | Transferenvironments | |
| C-maze | S-maze | |
| GAIL | -24.6±7.2 | 1.0±1.3 |
| VAIL | -65.6±18.9 | 20.8±39.7 |
| AIRL | -15.3±7.8 | -0.2±0.1 |
| AIRL - GP | -9.14±0.4 | -0.14±0.3 |
| VAIRL (β = 0) | -25.5±7.2 | 62.3±33.2 |
| VAIRL (ours) | -10.0±2.2 | 74.0±38.7 |
| VAIRL - GP (ours) | -9.18±0.4 | 156.5±5.6 |
| TRPO expert | -5.1 | 153.2 |
| Layer | Output size | Filter |
| FCReshape | 256·4·4256×4×4 | 256→256·4·4 |
| Resnet-blockUpsample | 128×4×4128×8×8 | 256→128 |
| Resnet-blockUpsample | 64×8×864 × 16 × 16 | 128→64 |
| Resnet-blockUpsample | 32 ×16×1632× 32×32 | 64→32 |
| Resnet-blockConv2D | 32×32×323× 32× 32 | 32→32 |
| Layer | Output size | Filter |
| Conv2D | 32 × 32 × 32 | 3→32 |
| Resnet-blockAvgPool2D | 64×32×3264 ×16 ×16 | 32→64 |
| Resnet-blockAvgPool2D | 128×16×16128×8×8 | 64→128 |
| Resnet-blockAvgPool2D | 256×8×8256×4×4 | 128→256 |
| FC | 1 | 256·4·4→1 |
| Layer | Output size | Filter |
| Conv2D | 32 ×32×32 | 3→32 |
| Resnet-blockAvgPool2D | 64×32×3264×16×16 | 32→64 |
| Resnet-blockAvgPool2D | 128×16×16128×8×8 | 64→128 |
| Resnet-blockAvgPool2D | 256×8×8256×4×4 | 128→256 |
| 1×1Conv2D | 2:256×4×4 | 256→2:256 |
| Sampling | 256×4×4 | |
| FC | 1 | 256·4·4→1 |
| Method | Backflip | Cartwheel | Dance | Run | Spinkick |
| BC | 3.01 | 2.88 | 2.93 | 2.63 | 2.88 |
| Merel et al., 2017 | 1.33 ± 0.03 | 1.47 ± 0.12 | 2.61 ± 0.30 | 0.52 ± 0.04 | 1.82 ± 0.35 |
| GAIL | 0.74±0.15 | 0.84± 0.05 | 1.31 ± 0.16 | 0.17±0.03 | 1.07 ± 0.03 |
| GAIL -noise | 0.42±0.02 | 0.92±0.07 | 0.96±0.08 | 0.21±0.05 | 0.95±0.14 |
| GAIL - noise z | 0.67±0.12 | 0.72± 0.04 | 1.14 ± 0.08 | 0.14±0.03 | 0.64±0.09 |
| GAIL - GP | 0.62±0.09 | 0.69 ±0.05 | 0.80± 0.32 | 0.12 ± 0.02 | 0.64± 0.04 |
| VAIL (ours) | 0.36±0.13 | 0.40±0.08 | 0.40±0.21 | 0.13±0.01 | 0.34± 0.05 |
| VAIL - GP (ours) | 0.46±0.17 | 0.31 ± 0.02 | 0.15±0.01 | 0.10±0.01 | 0.31 ± 0.02 |
| Peng et al., 2018 | 0.26 | 0.21 | 0.20 | 0.14 | 0.19 |
| Method | Transferenvironments | |
| C-maze | S-maze | |
| GAIL | -24.6±7.2 | 1.0±1.3 |
| VAIL | -65.6±18.9 | 20.8±39.7 |
| AIRL | -15.3±7.8 | -0.2±0.1 |
| AIRL - GP | -9.14±0.4 | -0.14±0.3 |
| VAIRL (β = 0) | -25.5±7.2 | 62.3±33.2 |
| VAIRL (ours) | -10.0±2.2 | 74.0±38.7 |
| VAIRL - GP (ours) | -9.18±0.4 | 156.5±5.6 |
| TRPO expert | -5.1 | 153.2 |
| Layer | Output size | Filter |
| FCReshape | 256·4·4256×4×4 | 256→256·4·4 |
| Resnet-blockUpsample | 128×4×4128×8×8 | 256→128 |
| Resnet-blockUpsample | 64×8×864 × 16 × 16 | 128→64 |
| Resnet-blockUpsample | 32 ×16×1632× 32×32 | 64→32 |
| Resnet-blockConv2D | 32×32×323× 32× 32 | 32→32 |
| Layer | Output size | Filter |
| Conv2D | 32 × 32 × 32 | 3→32 |
| Resnet-blockAvgPool2D | 64×32×3264 ×16 ×16 | 32→64 |
| Resnet-blockAvgPool2D | 128×16×16128×8×8 | 64→128 |
| Resnet-blockAvgPool2D | 256×8×8256×4×4 | 128→256 |
| FC | 1 | 256·4·4→1 |
| Layer | Output size | Filter |
| Conv2D | 32 ×32×32 | 3→32 |
| Resnet-blockAvgPool2D | 64×32×3264×16×16 | 32→64 |
| Resnet-blockAvgPool2D | 128×16×16128×8×8 | 64→128 |
| Resnet-blockAvgPool2D | 256×8×8256×4×4 | 128→256 |
| 1×1Conv2D | 2:256×4×4 | 256→2:256 |
| Sampling | 256×4×4 | |
| FC | 1 | 256·4·4→1 |
| Method | Backflip | Cartwheel | Dance | Run | Spinkick |
| BC | 3.01 | 2.88 | 2.93 | 2.63 | 2.88 |
| Merel et al., 2017 | 1.33 ± 0.03 | 1.47 ± 0.12 | 2.61 ± 0.30 | 0.52 ± 0.04 | 1.82 ± 0.35 |
| GAIL | 0.74±0.15 | 0.84± 0.05 | 1.31 ± 0.16 | 0.17±0.03 | 1.07 ± 0.03 |
| GAIL -noise | 0.42±0.02 | 0.92±0.07 | 0.96±0.08 | 0.21±0.05 | 0.95±0.14 |
| GAIL - noise z | 0.67±0.12 | 0.72± 0.04 | 1.14 ± 0.08 | 0.14±0.03 | 0.64±0.09 |
| GAIL - GP | 0.62±0.09 | 0.69 ±0.05 | 0.80± 0.32 | 0.12 ± 0.02 | 0.64± 0.04 |
| VAIL (ours) | 0.36±0.13 | 0.40±0.08 | 0.40±0.21 | 0.13±0.01 | 0.34± 0.05 |
| VAIL - GP (ours) | 0.46±0.17 | 0.31 ± 0.02 | 0.15±0.01 | 0.10±0.01 | 0.31 ± 0.02 |
| Peng et al., 2018 | 0.26 | 0.21 | 0.20 | 0.14 | 0.19 |
| Method | Backflip | Cartwheel | Dance | Run | Spinkick |
| BC | 3.01 | 2.88 | 2.93 | 2.63 | 2.88 |
| Merel et al., 2017 | 1.33 ± 0.03 | 1.47 ± 0.12 | 2.61 ± 0.30 | 0.52 ± 0.04 | 1.82 ± 0.35 |
| GAIL | 0.74±0.15 | 0.84± 0.05 | 1.31 ± 0.16 | 0.17±0.03 | 1.07 ± 0.03 |
| GAIL -noise | 0.42±0.02 | 0.92±0.07 | 0.96±0.08 | 0.21±0.05 | 0.95±0.14 |
| GAIL - noise z | 0.67±0.12 | 0.72± 0.04 | 1.14 ± 0.08 | 0.14±0.03 | 0.64±0.09 |
| GAIL - GP | 0.62±0.09 | 0.69 ±0.05 | 0.80± 0.32 | 0.12 ± 0.02 | 0.64± 0.04 |
| VAIL (ours) | 0.36±0.13 | 0.40±0.08 | 0.40±0.21 | 0.13±0.01 | 0.34± 0.05 |
| VAIL - GP (ours) | 0.46±0.17 | 0.31 ± 0.02 | 0.15±0.01 | 0.10±0.01 | 0.31 ± 0.02 |
| Peng et al., 2018 | 0.26 | 0.21 | 0.20 | 0.14 | 0.19 |
| Method | Transferenvironments | |
| C-maze | S-maze | |
| GAIL | -24.6±7.2 | 1.0±1.3 |
| VAIL | -65.6±18.9 | 20.8±39.7 |
| AIRL | -15.3±7.8 | -0.2±0.1 |
| AIRL - GP | -9.14±0.4 | -0.14±0.3 |
| VAIRL (β = 0) | -25.5±7.2 | 62.3±33.2 |
| VAIRL (ours) | -10.0±2.2 | 74.0±38.7 |
| VAIRL - GP (ours) | -9.18±0.4 | 156.5±5.6 |
| TRPO expert | -5.1 | 153.2 |
| Method | Transferenvironments | |
| C-maze | S-maze | |
| GAIL | -24.6±7.2 | 1.0±1.3 |
| VAIL | -65.6±18.9 | 20.8±39.7 |
| AIRL | -15.3±7.8 | -0.2±0.1 |
| AIRL - GP | -9.14±0.4 | -0.14±0.3 |
| VAIRL (β = 0) | -25.5±7.2 | 62.3±33.2 |
| VAIRL (ours) | -10.0±2.2 | 74.0±38.7 |
| VAIRL - GP (ours) | -9.18±0.4 | 156.5±5.6 |
| TRPO expert | -5.1 | 153.2 |
| Layer | Output size | Filter |
| FCReshape | 256·4·4256×4×4 | 256→256·4·4 |
| Resnet-blockUpsample | 128×4×4128×8×8 | 256→128 |
| Resnet-blockUpsample | 64×8×864 × 16 × 16 | 128→64 |
| Resnet-blockUpsample | 32 ×16×1632× 32×32 | 64→32 |
| Resnet-blockConv2D | 32×32×323× 32× 32 | 32→32 |
| Layer | Output size | Filter |
| Conv2D | 32 × 32 × 32 | 3→32 |
| Resnet-blockAvgPool2D | 64×32×3264 ×16 ×16 | 32→64 |
| Resnet-blockAvgPool2D | 128×16×16128×8×8 | 64→128 |
| Resnet-blockAvgPool2D | 256×8×8256×4×4 | 128→256 |
| FC | 1 | 256·4·4→1 |
| Layer | Output size | Filter |
| Conv2D | 32 ×32×32 | 3→32 |
| Resnet-blockAvgPool2D | 64×32×3264×16×16 | 32→64 |
| Resnet-blockAvgPool2D | 128×16×16128×8×8 | 64→128 |
| Resnet-blockAvgPool2D | 256×8×8256×4×4 | 128→256 |
| 1×1Conv2D | 2:256×4×4 | 256→2:256 |
| Sampling | 256×4×4 | |
| FC | 1 | 256·4·4→1 |
| Layer | Output size | Filter |
| FCReshape | 256·4·4256×4×4 | 256→256·4·4 |
| Resnet-blockUpsample | 128×4×4128×8×8 | 256→128 |
| Resnet-blockUpsample | 64×8×864 × 16 × 16 | 128→64 |
| Resnet-blockUpsample | 32 ×16×1632× 32×32 | 64→32 |
| Resnet-blockConv2D | 32×32×323× 32× 32 | 32→32 |
| Layer | Output size | Filter |
| Conv2D | 32 × 32 × 32 | 3→32 |
| Resnet-blockAvgPool2D | 64×32×3264 ×16 ×16 | 32→64 |
| Resnet-blockAvgPool2D | 128×16×16128×8×8 | 64→128 |
| Resnet-blockAvgPool2D | 256×8×8256×4×4 | 128→256 |
| FC | 1 | 256·4·4→1 |
| Layer | Output size | Filter |
| Conv2D | 32 ×32×32 | 3→32 |
| Resnet-blockAvgPool2D | 64×32×3264×16×16 | 32→64 |
| Resnet-blockAvgPool2D | 128×16×16128×8×8 | 64→128 |
| Resnet-blockAvgPool2D | 256×8×8256×4×4 | 128→256 |
| 1×1Conv2D | 2:256×4×4 | 256→2:256 |
| Sampling | 256×4×4 | |
| FC | 1 | 256·4·4→1 |
| Method | Backflip | Cartwheel | Dance | Run | Spinkick |
| BC | 3.01 | 2.88 | 2.93 | 2.63 | 2.88 |
| Merel et al., 2017 | 1.33 ± 0.03 | 1.47 ± 0.12 | 2.61 ± 0.30 | 0.52 ± 0.04 | 1.82 ± 0.35 |
| GAIL | 0.74±0.15 | 0.84± 0.05 | 1.31 ± 0.16 | 0.17±0.03 | 1.07 ± 0.03 |
| GAIL -noise | 0.42±0.02 | 0.92±0.07 | 0.96±0.08 | 0.21±0.05 | 0.95±0.14 |
| GAIL - noise z | 0.67±0.12 | 0.72± 0.04 | 1.14 ± 0.08 | 0.14±0.03 | 0.64±0.09 |
| GAIL - GP | 0.62±0.09 | 0.69 ±0.05 | 0.80± 0.32 | 0.12 ± 0.02 | 0.64± 0.04 |
| VAIL (ours) | 0.36±0.13 | 0.40±0.08 | 0.40±0.21 | 0.13±0.01 | 0.34± 0.05 |
| VAIL - GP (ours) | 0.46±0.17 | 0.31 ± 0.02 | 0.15±0.01 | 0.10±0.01 | 0.31 ± 0.02 |
| Peng et al., 2018 | 0.26 | 0.21 | 0.20 | 0.14 | 0.19 |
| Method | Transferenvironments | |
| C-maze | S-maze | |
| GAIL | -24.6±7.2 | 1.0±1.3 |
| VAIL | -65.6±18.9 | 20.8±39.7 |
| AIRL | -15.3±7.8 | -0.2±0.1 |
| AIRL - GP | -9.14±0.4 | -0.14±0.3 |
| VAIRL (β = 0) | -25.5±7.2 | 62.3±33.2 |
| VAIRL (ours) | -10.0±2.2 | 74.0±38.7 |
| VAIRL - GP (ours) | -9.18±0.4 | 156.5±5.6 |
| TRPO expert | -5.1 | 153.2 |
| Layer | Output size | Filter |
| FCReshape | 256·4·4256×4×4 | 256→256·4·4 |
| Resnet-blockUpsample | 128×4×4128×8×8 | 256→128 |
| Resnet-blockUpsample | 64×8×864 × 16 × 16 | 128→64 |
| Resnet-blockUpsample | 32 ×16×1632× 32×32 | 64→32 |
| Resnet-blockConv2D | 32×32×323× 32× 32 | 32→32 |
| Layer | Output size | Filter |
| Conv2D | 32 ×32×32 | 3→32 |
| Resnet-blockAvgPool2D | 64×32×3264×16×16 | 32→64 |
| Resnet-blockAvgPool2D | 128×16×16128×8×8 | 64→128 |
| Resnet-blockAvgPool2D | 256×8×8256×4×4 | 128→256 |
| 1×1Conv2D | 2:256×4×4 | 256→2:256 |
| Sampling | 256×4×4 | |
| FC | 1 | 256·4·4→1 |
| Layer | Output size | Filter |
| Conv2D | 32 × 32 × 32 | 3→32 |
| Resnet-blockAvgPool2D | 64×32×3264 ×16 ×16 | 32→64 |
| Resnet-blockAvgPool2D | 128×16×16128×8×8 | 64→128 |
| Resnet-blockAvgPool2D | 256×8×8256×4×4 | 128→256 |
| FC | 1 | 256·4·4→1 |
| g=15 | g=25 | g=50 | |
| MTMKL | 34.29 | 31.88 | 28.08 |
| GSC | 32.68 | 31.10 | 28.02 |
| VAR-BSC | 32.25 | 31.15 | 28.62 |
| TVAE | 34.27 ± .02 | 32.65 ± .06 | 29.61± .02 |
| g=15 | g=25 | g=50 | |
| N2v* | 32.05 | 29.20 | 25.42 |
| MTMKL | 34.29 | 31.88 | 28.08 |
| GSC | 33.78 | 32.01 | 28.35 |
| S5C | 33.50 | 32.08 | 28.35 |
| VAR-BSC | 33.50 | 32.32 | 28.91 |
| TVAE | 34.27± .02 | 32.65±.06 | 29.98 ± .05 |
| N2vt | 33.91 | 32.10 | 28.94 |
| KSVD | 34.32 | 32.15 | 27.95 |
| WNNM | 35.13 | 33.22 | 30.33 |
| BM3D | 34.94 | 32.86 | 29.37 |
| EPLL | 34.17 | 32.17 | 29.12 |
| BDGAN | 34.57 | 33.28 | 30.61 |
| DPDNN | 35.40 | 33.54 | 31.04 |
| Neural network units Input (H) 64 (512) Middle 64 (512) |
| Output (D) 64 (144) Cyclic Learning Rates |
| Min 1.r. 0.0001 |
| Max l.r. 0.01 (0.05) |
| Epochs/cycle 20 Batch size 32 |
| Evolutionaryparameters |
| Parents 10 (5) Children 9 (4) Generations 4(1) |
| 0=15 | g=25 | q=50 | |
| n2v* | 32.05 | 29.20 | 25.42 |
| n2vt | 32.93 | 32.10 | 20.96 |
| n2vt | 33.91 | 32.10 | 28.94 |
| TVAE | 34.27 ± .02 | 32.65 ± .06 | 29.98 ± .05 |
| g=15 | g=25 | g=50 | |
| MTMKL | 34.29 | 31.88 | 28.08 |
| GSC | 32.68 | 31.10 | 28.02 |
| VAR-BSC | 32.25 | 31.15 | 28.62 |
| TVAE | 34.27 ± .02 | 32.65 ± .06 | 29.61± .02 |
| g=15 | g=25 | g=50 | |
| N2v* | 32.05 | 29.20 | 25.42 |
| MTMKL | 34.29 | 31.88 | 28.08 |
| GSC | 33.78 | 32.01 | 28.35 |
| S5C | 33.50 | 32.08 | 28.35 |
| VAR-BSC | 33.50 | 32.32 | 28.91 |
| TVAE | 34.27± .02 | 32.65±.06 | 29.98 ± .05 |
| N2vt | 33.91 | 32.10 | 28.94 |
| KSVD | 34.32 | 32.15 | 27.95 |
| WNNM | 35.13 | 33.22 | 30.33 |
| BM3D | 34.94 | 32.86 | 29.37 |
| EPLL | 34.17 | 32.17 | 29.12 |
| BDGAN | 34.57 | 33.28 | 30.61 |
| DPDNN | 35.40 | 33.54 | 31.04 |
| Neural network units Input (H) 64 (512) Middle 64 (512) |
| Output (D) 64 (144) Cyclic Learning Rates |
| Min 1.r. 0.0001 |
| Max l.r. 0.01 (0.05) |
| Epochs/cycle 20 Batch size 32 |
| Evolutionaryparameters |
| Parents 10 (5) Children 9 (4) Generations 4(1) |
| 0=15 | g=25 | q=50 | |
| n2v* | 32.05 | 29.20 | 25.42 |
| n2vt | 32.93 | 32.10 | 20.96 |
| n2vt | 33.91 | 32.10 | 28.94 |
| TVAE | 34.27 ± .02 | 32.65 ± .06 | 29.98 ± .05 |
| g=15 | g=25 | g=50 | |
| MTMKL | 34.29 | 31.88 | 28.08 |
| GSC | 32.68 | 31.10 | 28.02 |
| VAR-BSC | 32.25 | 31.15 | 28.62 |
| TVAE | 34.27 ± .02 | 32.65 ± .06 | 29.61± .02 |
| g=15 | g=25 | g=50 | |
| MTMKL | 34.29 | 31.88 | 28.08 |
| GSC | 32.68 | 31.10 | 28.02 |
| VAR-BSC | 32.25 | 31.15 | 28.62 |
| TVAE | 34.27 ± .02 | 32.65 ± .06 | 29.61± .02 |
| g=15 | g=25 | g=50 | |
| N2v* | 32.05 | 29.20 | 25.42 |
| MTMKL | 34.29 | 31.88 | 28.08 |
| GSC | 33.78 | 32.01 | 28.35 |
| S5C | 33.50 | 32.08 | 28.35 |
| VAR-BSC | 33.50 | 32.32 | 28.91 |
| TVAE | 34.27± .02 | 32.65±.06 | 29.98 ± .05 |
| N2vt | 33.91 | 32.10 | 28.94 |
| KSVD | 34.32 | 32.15 | 27.95 |
| WNNM | 35.13 | 33.22 | 30.33 |
| BM3D | 34.94 | 32.86 | 29.37 |
| EPLL | 34.17 | 32.17 | 29.12 |
| BDGAN | 34.57 | 33.28 | 30.61 |
| DPDNN | 35.40 | 33.54 | 31.04 |
| g=15 | g=25 | g=50 | |
| N2v* | 32.05 | 29.20 | 25.42 |
| MTMKL | 34.29 | 31.88 | 28.08 |
| GSC | 33.78 | 32.01 | 28.35 |
| S5C | 33.50 | 32.08 | 28.35 |
| VAR-BSC | 33.50 | 32.32 | 28.91 |
| TVAE | 34.27± .02 | 32.65±.06 | 29.98 ± .05 |
| N2vt | 33.91 | 32.10 | 28.94 |
| KSVD | 34.32 | 32.15 | 27.95 |
| WNNM | 35.13 | 33.22 | 30.33 |
| BM3D | 34.94 | 32.86 | 29.37 |
| EPLL | 34.17 | 32.17 | 29.12 |
| BDGAN | 34.57 | 33.28 | 30.61 |
| DPDNN | 35.40 | 33.54 | 31.04 |
| Neural network units Input (H) 64 (512) Middle 64 (512) |
| Output (D) 64 (144) Cyclic Learning Rates |
| Min 1.r. 0.0001 |
| Max l.r. 0.01 (0.05) |
| Epochs/cycle 20 Batch size 32 |
| Evolutionaryparameters |
| Parents 10 (5) Children 9 (4) Generations 4(1) |
| 0=15 | g=25 | q=50 | |
| n2v* | 32.05 | 29.20 | 25.42 |
| n2vt | 32.93 | 32.10 | 20.96 |
| n2vt | 33.91 | 32.10 | 28.94 |
| TVAE | 34.27 ± .02 | 32.65 ± .06 | 29.98 ± .05 |
| Neural network units Input (H) 64 (512) Middle 64 (512) |
| Output (D) 64 (144) Cyclic Learning Rates |
| Min 1.r. 0.0001 |
| Max l.r. 0.01 (0.05) |
| Epochs/cycle 20 Batch size 32 |
| Evolutionaryparameters |
| Parents 10 (5) Children 9 (4) Generations 4(1) |
| 0=15 | g=25 | q=50 | |
| n2v* | 32.05 | 29.20 | 25.42 |
| n2vt | 32.93 | 32.10 | 20.96 |
| n2vt | 33.91 | 32.10 | 28.94 |
| TVAE | 34.27 ± .02 | 32.65 ± .06 | 29.98 ± .05 |
| Initialize model parameters Θ= {W,,²} Initialize eachΦ (n) with S distinct latent states repeat forall batches in dataset do forsample n in batch do (n) Φnew forall generations do |
| new = mutation (crossover (selection (Φnew))) Φ (n) Φ (n) UΦnew |
| end for (n) to S fitest elements based on (9) |
| TruncateΦ end for Use Adam to update W using objective (4) |
| g=15 | g=25 | g=50 | |
| MTMKL | 34.29 | 31.88 | 28.08 |
| GSC | 32.68 | 31.10 | 28.02 |
| VAR-BSC | 32.25 | 31.15 | 28.62 |
| TVAE | 34.27 ± .02 | 32.65 ± .06 | 29.61± .02 |
| g=15 | g=25 | g=50 | |
| N2v* | 32.05 | 29.20 | 25.42 |
| MTMKL | 34.29 | 31.88 | 28.08 |
| GSC | 33.78 | 32.01 | 28.35 |
| S5C | 33.50 | 32.08 | 28.35 |
| VAR-BSC | 33.50 | 32.32 | 28.91 |
| TVAE | 34.27± .02 | 32.65±.06 | 29.98 ± .05 |
| N2vt | 33.91 | 32.10 | 28.94 |
| KSVD | 34.32 | 32.15 | 27.95 |
| WNNM | 35.13 | 33.22 | 30.33 |
| BM3D | 34.94 | 32.86 | 29.37 |
| EPLL | 34.17 | 32.17 | 29.12 |
| BDGAN | 34.57 | 33.28 | 30.61 |
| DPDNN | 35.40 | 33.54 | 31.04 |
| 0=15 | g=25 | q=50 | |
| n2v* | 32.05 | 29.20 | 25.42 |
| n2vt | 32.93 | 32.10 | 20.96 |
| n2vt | 33.91 | 32.10 | 28.94 |
| TVAE | 34.27 ± .02 | 32.65 ± .06 | 29.98 ± .05 |
| Neural network units Input (H) 64 (512) Middle 64 (512) |
| Output (D) 64 (144) Cyclic Learning Rates |
| Min 1.r. 0.0001 |
| Max l.r. 0.01 (0.05) |
| Epochs/cycle 20 Batch size 32 |
| Evolutionaryparameters |
| Parents 10 (5) Children 9 (4) Generations 4(1) |
| Module | Input | Output | Description |
| find | P,H | Hent | Forrelated entities in question,select the relevant tokens from dialogue history |
| summarize | Hent,Q | Qctx | Based on contextual entity representations,summarise the question semantics |
| where | P,V | Vent | Select the relevant spatial position corresponding to original (resolved) entities |
| when | P,Vent | Vent+act | Select the relevant entity-aware temporal steps corresponding to the action parameter |
| describe | P,Vent+act | Vctx Vctx | Select visual entity-action features based on non-binary question types |
| exist | Q,Vent+act | Select visual entity-action features based on binary (yes/no) question types |
| Model | PT | Visual | Audio | BLEU4 | METEOR | ROUGE-L | CIDEr |
| Audio/Visual only (without Video Summary/Caption) | |||||||
| Baseline (Hori et al.,2019) | I | ■ | 0.305 | 0.217 | 0.481 | 0.733 | |
| Baseline (Hori et al.,2019) | I | V | 0.309 | 0.215 | 0.487 | 0.746 | |
| Baseline+GRU+Attn.(Le et al.,2019a) | I | V | 0.315 | 0.239 | 0.509 | 0.848 | |
| FGA (Schwartz et al.,2019) | I | V | 0.806 | ||||
| JMAN (Chu et al., 2020) | I | - | 0.309 | 0.240 | 0.520 | 0.890 | |
| Student-Teacher (Hori et al.,2019) | I | V | 0.371 | 0.248 | 0.527 | 0.966 | |
| MTN (Le et al.,2019b) | I | - | 0.343 | 0.247 | 0.520 | 0.936 | |
| MTN (Le et al., 2019b) | I | V | 0.368 | 0.259 | 0.537 | 0.964 | |
| MSTN (Lee et al.,2020) | I | V | 0.379 | 0.261 | 0.548 | 1.028 | |
| RLM-GPT2 (Li et al., 2020) | √ | I | V | 0.402 | 0.254 | 0.544 | 1.052 |
| VilNMN | = | I | - | 0.397 | 0.262 | 0.550 | 1.059 |
| VilNMN | FR | ■ | 0.388 | 0.259 | 0.549 | 1.040 | |
| VilNMN | - | V | 0.381 | 0.252 | 0.534 | 1.004 | |
| VilNMN | I | V | 0.396 | 0.263 | 0.549 | 1.059 | |
| Audio/Visual only (with Video Summary/Caption) | |||||||
| TopicEmb (Kumar et al.,2019) | I | A | 0.329 | 0.223 | 0.488 | 0.762 | |
| Baseline+GRU+Attn.(Le et al.,2019a) | I | V | 0.310 | 0.242 | 0.515 | 0.856 | |
| JMAN (Chu et al., 2020) | I | - | 0.334 | 0.239 | 0.533 | 0.941 | |
| FA+HRED (Nguyen et al., 2018) | = | I | V | 0.360 | 0.249 | 0.544 | 0.997 |
| VideoSum (Sanabria etal.,2019) | = | RX | - | 0.394 | 0.267 | 0.563 | 1.094 |
| VideoSum+How2 (Sanabria et al.,2019) | √ | RX | - | 0.387 | 0.266 | 0.564 | 1.087 |
| MSTN (Lee et al.,2020) | I | V | 0.377 | 0.275 | 0.566 | 1.115 | |
| Student-Teacher (Hori et al.,2019) | I | V | 0.405 | 0.273 | 0.566 | 1.118 | |
| MTN (Le et al.,2019b) | I | - | 0.392 | 0.269 | 0.559 | 1.066 | |
| MTN (Le et al.,2019b) | I | V | 0.410 | 0.274 | 0.569 | 1.129 | |
| VGD-GPT2 (Le & Hoi,2020) | √ | I | V | 0.436 | 0.282 | 0.579 | 1.194 |
| RLM-GPT2 (Li et al., 2020) | √ | I | V | 0.459 | 0.294 | 0.606 | 1.308 |
| VilNMN | ■ | I | 1 | 0.421 | 0.277 | 0.574 | 1.171 |
| VilNMN | FR | - | 0.421 | 0.275 | 0.571 | 1.148 | |
| VilNMN | I | V | 0.421 | 0.277 | 0.573 | 1.167 | |
| VilNMN | I+C | V | 0.429 | 0.278 | 0.578 | 1.188 | |
| # | ModelVariant | BLEU4 | CIDEr |
| A | FullVilNMN | 0.421 | 1.171 |
| B | No video NMNs; + vanilla text-→video attention | 0.415 | 1.159 |
| C | →No dial.NMNs;+ response-→history attention | 0.412 | 1.151 |
| D | →No dial.NMNs;+ response-→concat(history+question) attention | 0.411 | 1.133 |
| E | No dial.NMNs;+ HREDLsTM(history) + question attn. | 0.414 | 1.153 |
| F | →No dial.NMNs; + HREDGRU(history) + question attn. | 0.415 | 1.138 |
| Turn | Question | Dialogue NMN | Video NMN | Response |
| 3 | ...what are theydoing in the scene ？are they on arooftop？ | Predicted:find(they)，find(the scene)→summarize()×Gold: find(they)→summarize() | Predicted: where(two men in thevideo),where(the scene)→when(doing inthe scene)→describe(what)Gold:where(two men)，where (rooftop)→when(doing in the scene)-describe(what) | Predicted: one is washinga chair and the other istaking picturesXGold: yes,on a secondfloor roof deck ,one man iswashing a chair,anotherman is either filming ortaking pictures |
| 4 | does he sit in thechair after washingit？ | Predicted:find (he),find(the chair)→summarize()Gold: find(he),find(the chair)→summarize() | Predicted: where(one man in the video)，where(a chair)→ when(sit in the chairafter washing it)-exist()Gold:where(one man)，where(a chair)→when(sit in the chair after washingit)-exist() | Predicted: no, he does notsit in the chair√Gold: no he does not,thereis a pipe with water runningall over |
| Model | Visual | Count (Loss) | Action (Acc) | Transition (Acc) | FrameQA (Acc) |
| VIS(avg) (Ren et al.,2015a) | R | 4.80 | 0.488 | 0.348 | 0.350 |
| MCB (aggr) (Fukui et al.,2016) | R | 5.17 | 0.589 | 0.243 | 0.257 |
| Yu et al. (Yu et al., 2017) | R | 5.13 | 0.561 | 0.640 | 0.396 |
| ST-VQA (t) (Gao et al., 2018) | R+F | 4.32 | 0.629 | 0.694 | 0.495 |
| Co-Mem (Gao et al.,2018) | R+F | 4.10 | 0.682 | 0.743 | 0.515 |
| PSAC (Li et al.,2019) | R | 4.27 | 0.704 | 0.769 | 0.557 |
| HME (Fan et al.,2019) | R+C | 4.02 | 0.739 | 0.778 | 0.538 |
| STA (Gao et al.,2019) | R | 4.25 | 0.723 | 0.790 | 0.566 |
| CRN+MAC (Le et al.,2019c) | R | 4.23 | 0.713 | 0.787 | 0.592 |
| MQL (Lei et al., 2020) | V | - | - | - | 0.598 |
| QueST (Jiang et al.,2020) | R | 4.19 | 0.759 | 0.810 | 0.597 |
| HGA (Jiang & Han,2020) | R+C | 4.09 | 0.754 | 0.810 | 0.551 |
| GCN (Huang et al., 2020) | R+C | 3.95 | 0.743 | 0.811 | 0.563 |
| HCRN (Le et al.,2020) | R+RX | 3.82 | 0.750 | 0.814 | 0.559 |
| VilNMN | R | 2.65 | 0.845 | 0.887 | 0.747 |
| →soft label programs | R | 1.90 | 0.857 | 0.898 | 0.780 |
| →- res-to-question attn. | R | 3.28 | 0.801 | 0.776 | 0.679 |
| # | Train | Val. | Test | |
| AVSD | Dialogs | 7,659 | 1,787 | 1,710 |
| Turns | 153,180 | 35,740 | 13,490 | |
| Words | 1,450,754 | 339,006 | 110,252 | |
| TGIFQA | Count QA | 24,159 | 2,684 | 3,554 |
| Action QA | 18,428 | 2,047 | 2,274 | |
| Trans. QA | 47,434 | 5,270 | 6,232 | |
| Frame QA | 35,453 | 3,939 | 13,691 | |
| BLEU4 | CIDEr | |||
| turn position | Model B | Model E | Model B | Model E |
| 1 | 0.579 | 0.587 | 1.623 | 1.650 |
| 2 | 0.429 | 0.430 | 1.155 | 1.142 |
| 3 | 0.275 | 0.289 | 0.867 | 0.846 |
| 4 | 0.309 | 0.305 | 0.859 | 0.855 |
| 5 | 0.355 | 0.335 | 1.088 | 1.023 |
| 6 | 0.357 | 0.329 | 1.044 | 0.950 |
| 7 | 0.342 | 0.325 | 0.896 | 0.847 |
| 8 | 0.361 | 0.332 | 1.025 | 0.973 |
| 9 | 0.383 | 0.431 | 1.043 | 1.182 |
| 10 | 0.395 | 0.371 | 0.931 | 0.977 |
| BLEU4 | CIDEr | |||
| video range (seconds) | Model A | Model B | Model A | Model B |
| 1-23 | 0.432 | 0.447 | 1.298 | 1.355 |
| 23-28 | 0.436 | 0.433 | 1.264 | 1.165 |
| 28-30 | 0.398 | 0.376 | 1.203 | 1.164 |
| 30-30.6 | 0.441 | 0.418 | 1.220 | 1.202 |
| 30.6-31 | 0.413 | 0.411 | 1.250 | 1.166 |
| 31-31.6 | 0.439 | 0.451 | 1.249 | 1.295 |
| 31.6-32 | 0.430 | 0.419 | 1.217 | 1.192 |
| 32-33 | 0.468 | 0.445 | 1.343 | 1.237 |
| 33-37 | 0.388 | 0.381 | 1.149 | 1.124 |
| 37-75 | 0.356 | 0.365 | 0.910 | 0.962 |
| Turn | Question | Dialogue UnderstandingProgram | Video UnderstandingProgram | Response |
| 1 | hi how many peopleare in the video fromstart to finish ? | Predicted: summari ze()Gold: summarize () | Predicted: describe (how manypeople)Gold:describe (how manypeople) | Predicted: there are twopeople in the video.Gold: hello,there aretwo men |
| 2 | are they carrying ona conversationbetween themselves？ | Predicted:find(they)→summarize()Gold:find(they)，find(themselves)→summarize() | Predicted: where(two men in the←video)→when(carrying on aconversation betweenthemselves)-exist()Gold:where(two men)→ when(carrying a conversationbetween themselves)-exist() | Predicted: yes,they arecarrying a conversation×Gold: no,backgroundnoise is birdsong andshuffling noises |
| 3 | what are they doingin the scene ?arethey on a rooftop ? | Predicted: find(they)，find(the scene)→summarize() ×Gold: find(they)→summarize() | Predicted:where(two men in thevideo),where(the scene)→when(doing in the scene)→describe(what) ×Gold: where(two men),where(rooftop)→when(doingin the scene)-describe(what) | Predicted: one iswashing a chair and theother is taking pictures×Gold: yes,on a secondfloor roof deck ,oneman is washing a chair,another man is eitherfilming or taking pictures |
| 4 | does he sit in thechair after washingit？ | Predicted: find(he),find(the chair)→summarize()Gold: find (he),find(the chair)→summarize() | Predicted: where(one man in thevideo)，where(a chair)→when(sit in the chair afterwashing it)→exist()Gold:where(one man)，where(achair)→ when(sit in thechair after washing it)→exist() | Predicted: no,he doesnot sit in the chairGold: no he does not,there is a pipe withwater running all over |
| Module | Input | Output | Description |
| find | P,H | Hent | Forrelated entities in question,select the relevant tokens from dialogue history |
| summarize | Hent,Q | Qctx | Based on contextual entity representations,summarise the question semantics |
| where | P,V | Vent | Select the relevant spatial position corresponding to original (resolved) entities |
| when | P,Vent | Vent+act | Select the relevant entity-aware temporal steps corresponding to the action parameter |
| describe | P,Vent+act | Vctx Vctx | Select visual entity-action features based on non-binary question types |
| exist | Q,Vent+act | Select visual entity-action features based on binary (yes/no) question types |
| Model | PT | Visual | Audio | BLEU4 | METEOR | ROUGE-L | CIDEr |
| Audio/Visual only (without Video Summary/Caption) | |||||||
| Baseline (Hori et al.,2019) | I | ■ | 0.305 | 0.217 | 0.481 | 0.733 | |
| Baseline (Hori et al.,2019) | I | V | 0.309 | 0.215 | 0.487 | 0.746 | |
| Baseline+GRU+Attn.(Le et al.,2019a) | I | V | 0.315 | 0.239 | 0.509 | 0.848 | |
| FGA (Schwartz et al.,2019) | I | V | 0.806 | ||||
| JMAN (Chu et al., 2020) | I | - | 0.309 | 0.240 | 0.520 | 0.890 | |
| Student-Teacher (Hori et al.,2019) | I | V | 0.371 | 0.248 | 0.527 | 0.966 | |
| MTN (Le et al.,2019b) | I | - | 0.343 | 0.247 | 0.520 | 0.936 | |
| MTN (Le et al., 2019b) | I | V | 0.368 | 0.259 | 0.537 | 0.964 | |
| MSTN (Lee et al.,2020) | I | V | 0.379 | 0.261 | 0.548 | 1.028 | |
| RLM-GPT2 (Li et al., 2020) | √ | I | V | 0.402 | 0.254 | 0.544 | 1.052 |
| VilNMN | = | I | - | 0.397 | 0.262 | 0.550 | 1.059 |
| VilNMN | FR | ■ | 0.388 | 0.259 | 0.549 | 1.040 | |
| VilNMN | - | V | 0.381 | 0.252 | 0.534 | 1.004 | |
| VilNMN | I | V | 0.396 | 0.263 | 0.549 | 1.059 | |
| Audio/Visual only (with Video Summary/Caption) | |||||||
| TopicEmb (Kumar et al.,2019) | I | A | 0.329 | 0.223 | 0.488 | 0.762 | |
| Baseline+GRU+Attn.(Le et al.,2019a) | I | V | 0.310 | 0.242 | 0.515 | 0.856 | |
| JMAN (Chu et al., 2020) | I | - | 0.334 | 0.239 | 0.533 | 0.941 | |
| FA+HRED (Nguyen et al., 2018) | = | I | V | 0.360 | 0.249 | 0.544 | 0.997 |
| VideoSum (Sanabria etal.,2019) | = | RX | - | 0.394 | 0.267 | 0.563 | 1.094 |
| VideoSum+How2 (Sanabria et al.,2019) | √ | RX | - | 0.387 | 0.266 | 0.564 | 1.087 |
| MSTN (Lee et al.,2020) | I | V | 0.377 | 0.275 | 0.566 | 1.115 | |
| Student-Teacher (Hori et al.,2019) | I | V | 0.405 | 0.273 | 0.566 | 1.118 | |
| MTN (Le et al.,2019b) | I | - | 0.392 | 0.269 | 0.559 | 1.066 | |
| MTN (Le et al.,2019b) | I | V | 0.410 | 0.274 | 0.569 | 1.129 | |
| VGD-GPT2 (Le & Hoi,2020) | √ | I | V | 0.436 | 0.282 | 0.579 | 1.194 |
| RLM-GPT2 (Li et al., 2020) | √ | I | V | 0.459 | 0.294 | 0.606 | 1.308 |
| VilNMN | ■ | I | 1 | 0.421 | 0.277 | 0.574 | 1.171 |
| VilNMN | FR | - | 0.421 | 0.275 | 0.571 | 1.148 | |
| VilNMN | I | V | 0.421 | 0.277 | 0.573 | 1.167 | |
| VilNMN | I+C | V | 0.429 | 0.278 | 0.578 | 1.188 |
| # | ModelVariant | BLEU4 | CIDEr |
| A | FullVilNMN | 0.421 | 1.171 |
| B | No video NMNs; + vanilla text-→video attention | 0.415 | 1.159 |
| C | →No dial.NMNs;+ response-→history attention | 0.412 | 1.151 |
| D | →No dial.NMNs;+ response-→concat(history+question) attention | 0.411 | 1.133 |
| E | No dial.NMNs;+ HREDLsTM(history) + question attn. | 0.414 | 1.153 |
| F | →No dial.NMNs; + HREDGRU(history) + question attn. | 0.415 | 1.138 |
| Turn | Question | Dialogue NMN | Video NMN | Response |
| 3 | ...what are theydoing in the scene ？are they on arooftop？ | Predicted:find(they)，find(the scene)→summarize()×Gold: find(they)→summarize() | Predicted: where(two men in thevideo),where(the scene)→when(doing inthe scene)→describe(what)Gold:where(two men)，where (rooftop)→when(doing in the scene)-describe(what) | Predicted: one is washinga chair and the other istaking picturesXGold: yes,on a secondfloor roof deck ,one man iswashing a chair,anotherman is either filming ortaking pictures |
| 4 | does he sit in thechair after washingit？ | Predicted:find (he),find(the chair)→summarize()Gold: find(he),find(the chair)→summarize() | Predicted: where(one man in the video)，where(a chair)→ when(sit in the chairafter washing it)-exist()Gold:where(one man)，where(a chair)→when(sit in the chair after washingit)-exist() | Predicted: no, he does notsit in the chair√Gold: no he does not,thereis a pipe with water runningall over |
| Model | Visual | Count (Loss) | Action (Acc) | Transition (Acc) | FrameQA (Acc) |
| VIS(avg) (Ren et al.,2015a) | R | 4.80 | 0.488 | 0.348 | 0.350 |
| MCB (aggr) (Fukui et al.,2016) | R | 5.17 | 0.589 | 0.243 | 0.257 |
| Yu et al. (Yu et al., 2017) | R | 5.13 | 0.561 | 0.640 | 0.396 |
| ST-VQA (t) (Gao et al., 2018) | R+F | 4.32 | 0.629 | 0.694 | 0.495 |
| Co-Mem (Gao et al.,2018) | R+F | 4.10 | 0.682 | 0.743 | 0.515 |
| PSAC (Li et al.,2019) | R | 4.27 | 0.704 | 0.769 | 0.557 |
| HME (Fan et al.,2019) | R+C | 4.02 | 0.739 | 0.778 | 0.538 |
| STA (Gao et al.,2019) | R | 4.25 | 0.723 | 0.790 | 0.566 |
| CRN+MAC (Le et al.,2019c) | R | 4.23 | 0.713 | 0.787 | 0.592 |
| MQL (Lei et al., 2020) | V | - | - | - | 0.598 |
| QueST (Jiang et al.,2020) | R | 4.19 | 0.759 | 0.810 | 0.597 |
| HGA (Jiang & Han,2020) | R+C | 4.09 | 0.754 | 0.810 | 0.551 |
| GCN (Huang et al., 2020) | R+C | 3.95 | 0.743 | 0.811 | 0.563 |
| HCRN (Le et al.,2020) | R+RX | 3.82 | 0.750 | 0.814 | 0.559 |
| VilNMN | R | 2.65 | 0.845 | 0.887 | 0.747 |
| →soft label programs | R | 1.90 | 0.857 | 0.898 | 0.780 |
| →- res-to-question attn. | R | 3.28 | 0.801 | 0.776 | 0.679 |
| # | Train | Val. | Test | |
| AVSD | Dialogs | 7,659 | 1,787 | 1,710 |
| Turns | 153,180 | 35,740 | 13,490 | |
| Words | 1,450,754 | 339,006 | 110,252 | |
| TGIFQA | Count QA | 24,159 | 2,684 | 3,554 |
| Action QA | 18,428 | 2,047 | 2,274 | |
| Trans. QA | 47,434 | 5,270 | 6,232 | |
| Frame QA | 35,453 | 3,939 | 13,691 |
| BLEU4 | CIDEr | |||
| turn position | Model B | Model E | Model B | Model E |
| 1 | 0.579 | 0.587 | 1.623 | 1.650 |
| 2 | 0.429 | 0.430 | 1.155 | 1.142 |
| 3 | 0.275 | 0.289 | 0.867 | 0.846 |
| 4 | 0.309 | 0.305 | 0.859 | 0.855 |
| 5 | 0.355 | 0.335 | 1.088 | 1.023 |
| 6 | 0.357 | 0.329 | 1.044 | 0.950 |
| 7 | 0.342 | 0.325 | 0.896 | 0.847 |
| 8 | 0.361 | 0.332 | 1.025 | 0.973 |
| 9 | 0.383 | 0.431 | 1.043 | 1.182 |
| 10 | 0.395 | 0.371 | 0.931 | 0.977 |
| BLEU4 | CIDEr | |||
| video range (seconds) | Model A | Model B | Model A | Model B |
| 1-23 | 0.432 | 0.447 | 1.298 | 1.355 |
| 23-28 | 0.436 | 0.433 | 1.264 | 1.165 |
| 28-30 | 0.398 | 0.376 | 1.203 | 1.164 |
| 30-30.6 | 0.441 | 0.418 | 1.220 | 1.202 |
| 30.6-31 | 0.413 | 0.411 | 1.250 | 1.166 |
| 31-31.6 | 0.439 | 0.451 | 1.249 | 1.295 |
| 31.6-32 | 0.430 | 0.419 | 1.217 | 1.192 |
| 32-33 | 0.468 | 0.445 | 1.343 | 1.237 |
| 33-37 | 0.388 | 0.381 | 1.149 | 1.124 |
| 37-75 | 0.356 | 0.365 | 0.910 | 0.962 |
| Turn | Question | Dialogue UnderstandingProgram | Video UnderstandingProgram | Response |
| 1 | hi how many peopleare in the video fromstart to finish ? | Predicted: summari ze()Gold: summarize () | Predicted: describe (how manypeople)Gold:describe (how manypeople) | Predicted: there are twopeople in the video.Gold: hello,there aretwo men |
| 2 | are they carrying ona conversationbetween themselves？ | Predicted:find(they)→summarize()Gold:find(they)，find(themselves)→summarize() | Predicted: where(two men in the←video)→when(carrying on aconversation betweenthemselves)-exist()Gold:where(two men)→ when(carrying a conversationbetween themselves)-exist() | Predicted: yes,they arecarrying a conversation×Gold: no,backgroundnoise is birdsong andshuffling noises |
| 3 | what are they doingin the scene ?arethey on a rooftop ? | Predicted: find(they)，find(the scene)→summarize() ×Gold: find(they)→summarize() | Predicted:where(two men in thevideo),where(the scene)→when(doing in the scene)→describe(what) ×Gold: where(two men),where(rooftop)→when(doingin the scene)-describe(what) | Predicted: one iswashing a chair and theother is taking pictures×Gold: yes,on a secondfloor roof deck ,oneman is washing a chair,another man is eitherfilming or taking pictures |
| 4 | does he sit in thechair after washingit？ | Predicted: find(he),find(the chair)→summarize()Gold: find (he),find(the chair)→summarize() | Predicted: where(one man in thevideo)，where(a chair)→when(sit in the chair afterwashing it)→exist()Gold:where(one man)，where(achair)→ when(sit in thechair after washing it)→exist() | Predicted: no,he doesnot sit in the chairGold: no he does not,there is a pipe withwater running all over |
| Module | Input | Output | Description |
| find | P,H | Hent | Forrelated entities in question,select the relevant tokens from dialogue history |
| summarize | Hent,Q | Qctx | Based on contextual entity representations,summarise the question semantics |
| where | P,V | Vent | Select the relevant spatial position corresponding to original (resolved) entities |
| when | P,Vent | Vent+act | Select the relevant entity-aware temporal steps corresponding to the action parameter |
| describe | P,Vent+act | Vctx Vctx | Select visual entity-action features based on non-binary question types |
| exist | Q,Vent+act | Select visual entity-action features based on binary (yes/no) question types |
| Module | Input | Output | Description |
| find | P,H | Hent | Forrelated entities in question,select the relevant tokens from dialogue history |
| summarize | Hent,Q | Qctx | Based on contextual entity representations,summarise the question semantics |
| where | P,V | Vent | Select the relevant spatial position corresponding to original (resolved) entities |
| when | P,Vent | Vent+act | Select the relevant entity-aware temporal steps corresponding to the action parameter |
| describe | P,Vent+act | Vctx Vctx | Select visual entity-action features based on non-binary question types |
| exist | Q,Vent+act | Select visual entity-action features based on binary (yes/no) question types |
| Model | PT | Visual | Audio | BLEU4 | METEOR | ROUGE-L | CIDEr |
| Audio/Visual only (without Video Summary/Caption) | |||||||
| Baseline (Hori et al.,2019) | I | ■ | 0.305 | 0.217 | 0.481 | 0.733 | |
| Baseline (Hori et al.,2019) | I | V | 0.309 | 0.215 | 0.487 | 0.746 | |
| Baseline+GRU+Attn.(Le et al.,2019a) | I | V | 0.315 | 0.239 | 0.509 | 0.848 | |
| FGA (Schwartz et al.,2019) | I | V | 0.806 | ||||
| JMAN (Chu et al., 2020) | I | - | 0.309 | 0.240 | 0.520 | 0.890 | |
| Student-Teacher (Hori et al.,2019) | I | V | 0.371 | 0.248 | 0.527 | 0.966 | |
| MTN (Le et al.,2019b) | I | - | 0.343 | 0.247 | 0.520 | 0.936 | |
| MTN (Le et al., 2019b) | I | V | 0.368 | 0.259 | 0.537 | 0.964 | |
| MSTN (Lee et al.,2020) | I | V | 0.379 | 0.261 | 0.548 | 1.028 | |
| RLM-GPT2 (Li et al., 2020) | √ | I | V | 0.402 | 0.254 | 0.544 | 1.052 |
| VilNMN | = | I | - | 0.397 | 0.262 | 0.550 | 1.059 |
| VilNMN | FR | ■ | 0.388 | 0.259 | 0.549 | 1.040 | |
| VilNMN | - | V | 0.381 | 0.252 | 0.534 | 1.004 | |
| VilNMN | I | V | 0.396 | 0.263 | 0.549 | 1.059 | |
| Audio/Visual only (with Video Summary/Caption) | |||||||
| TopicEmb (Kumar et al.,2019) | I | A | 0.329 | 0.223 | 0.488 | 0.762 | |
| Baseline+GRU+Attn.(Le et al.,2019a) | I | V | 0.310 | 0.242 | 0.515 | 0.856 | |
| JMAN (Chu et al., 2020) | I | - | 0.334 | 0.239 | 0.533 | 0.941 | |
| FA+HRED (Nguyen et al., 2018) | = | I | V | 0.360 | 0.249 | 0.544 | 0.997 |
| VideoSum (Sanabria etal.,2019) | = | RX | - | 0.394 | 0.267 | 0.563 | 1.094 |
| VideoSum+How2 (Sanabria et al.,2019) | √ | RX | - | 0.387 | 0.266 | 0.564 | 1.087 |
| MSTN (Lee et al.,2020) | I | V | 0.377 | 0.275 | 0.566 | 1.115 | |
| Student-Teacher (Hori et al.,2019) | I | V | 0.405 | 0.273 | 0.566 | 1.118 | |
| MTN (Le et al.,2019b) | I | - | 0.392 | 0.269 | 0.559 | 1.066 | |
| MTN (Le et al.,2019b) | I | V | 0.410 | 0.274 | 0.569 | 1.129 | |
| VGD-GPT2 (Le & Hoi,2020) | √ | I | V | 0.436 | 0.282 | 0.579 | 1.194 |
| RLM-GPT2 (Li et al., 2020) | √ | I | V | 0.459 | 0.294 | 0.606 | 1.308 |
| VilNMN | ■ | I | 1 | 0.421 | 0.277 | 0.574 | 1.171 |
| VilNMN | FR | - | 0.421 | 0.275 | 0.571 | 1.148 | |
| VilNMN | I | V | 0.421 | 0.277 | 0.573 | 1.167 | |
| VilNMN | I+C | V | 0.429 | 0.278 | 0.578 | 1.188 |
| # | ModelVariant | BLEU4 | CIDEr |
| A | FullVilNMN | 0.421 | 1.171 |
| B | No video NMNs; + vanilla text-→video attention | 0.415 | 1.159 |
| C | →No dial.NMNs;+ response-→history attention | 0.412 | 1.151 |
| D | →No dial.NMNs;+ response-→concat(history+question) attention | 0.411 | 1.133 |
| E | No dial.NMNs;+ HREDLsTM(history) + question attn. | 0.414 | 1.153 |
| F | →No dial.NMNs; + HREDGRU(history) + question attn. | 0.415 | 1.138 |
| Model | PT | Visual | Audio | BLEU4 | METEOR | ROUGE-L | CIDEr |
| Audio/Visual only (without Video Summary/Caption) | |||||||
| Baseline (Hori et al.,2019) | I | ■ | 0.305 | 0.217 | 0.481 | 0.733 | |
| Baseline (Hori et al.,2019) | I | V | 0.309 | 0.215 | 0.487 | 0.746 | |
| Baseline+GRU+Attn.(Le et al.,2019a) | I | V | 0.315 | 0.239 | 0.509 | 0.848 | |
| FGA (Schwartz et al.,2019) | I | V | 0.806 | ||||
| JMAN (Chu et al., 2020) | I | - | 0.309 | 0.240 | 0.520 | 0.890 | |
| Student-Teacher (Hori et al.,2019) | I | V | 0.371 | 0.248 | 0.527 | 0.966 | |
| MTN (Le et al.,2019b) | I | - | 0.343 | 0.247 | 0.520 | 0.936 | |
| MTN (Le et al., 2019b) | I | V | 0.368 | 0.259 | 0.537 | 0.964 | |
| MSTN (Lee et al.,2020) | I | V | 0.379 | 0.261 | 0.548 | 1.028 | |
| RLM-GPT2 (Li et al., 2020) | √ | I | V | 0.402 | 0.254 | 0.544 | 1.052 |
| VilNMN | = | I | - | 0.397 | 0.262 | 0.550 | 1.059 |
| VilNMN | FR | ■ | 0.388 | 0.259 | 0.549 | 1.040 | |
| VilNMN | - | V | 0.381 | 0.252 | 0.534 | 1.004 | |
| VilNMN | I | V | 0.396 | 0.263 | 0.549 | 1.059 | |
| Audio/Visual only (with Video Summary/Caption) | |||||||
| TopicEmb (Kumar et al.,2019) | I | A | 0.329 | 0.223 | 0.488 | 0.762 | |
| Baseline+GRU+Attn.(Le et al.,2019a) | I | V | 0.310 | 0.242 | 0.515 | 0.856 | |
| JMAN (Chu et al., 2020) | I | - | 0.334 | 0.239 | 0.533 | 0.941 | |
| FA+HRED (Nguyen et al., 2018) | = | I | V | 0.360 | 0.249 | 0.544 | 0.997 |
| VideoSum (Sanabria etal.,2019) | = | RX | - | 0.394 | 0.267 | 0.563 | 1.094 |
| VideoSum+How2 (Sanabria et al.,2019) | √ | RX | - | 0.387 | 0.266 | 0.564 | 1.087 |
| MSTN (Lee et al.,2020) | I | V | 0.377 | 0.275 | 0.566 | 1.115 | |
| Student-Teacher (Hori et al.,2019) | I | V | 0.405 | 0.273 | 0.566 | 1.118 | |
| MTN (Le et al.,2019b) | I | - | 0.392 | 0.269 | 0.559 | 1.066 | |
| MTN (Le et al.,2019b) | I | V | 0.410 | 0.274 | 0.569 | 1.129 | |
| VGD-GPT2 (Le & Hoi,2020) | √ | I | V | 0.436 | 0.282 | 0.579 | 1.194 |
| RLM-GPT2 (Li et al., 2020) | √ | I | V | 0.459 | 0.294 | 0.606 | 1.308 |
| VilNMN | ■ | I | 1 | 0.421 | 0.277 | 0.574 | 1.171 |
| VilNMN | FR | - | 0.421 | 0.275 | 0.571 | 1.148 | |
| VilNMN | I | V | 0.421 | 0.277 | 0.573 | 1.167 | |
| VilNMN | I+C | V | 0.429 | 0.278 | 0.578 | 1.188 |
| # | ModelVariant | BLEU4 | CIDEr |
| A | FullVilNMN | 0.421 | 1.171 |
| B | No video NMNs; + vanilla text-→video attention | 0.415 | 1.159 |
| C | →No dial.NMNs;+ response-→history attention | 0.412 | 1.151 |
| D | →No dial.NMNs;+ response-→concat(history+question) attention | 0.411 | 1.133 |
| E | No dial.NMNs;+ HREDLsTM(history) + question attn. | 0.414 | 1.153 |
| F | →No dial.NMNs; + HREDGRU(history) + question attn. | 0.415 | 1.138 |
| Turn | Question | Dialogue NMN | Video NMN | Response |
| 3 | ...what are theydoing in the scene ？are they on arooftop？ | Predicted:find(they)，find(the scene)→summarize()×Gold: find(they)→summarize() | Predicted: where(two men in thevideo),where(the scene)→when(doing inthe scene)→describe(what)Gold:where(two men)，where (rooftop)→when(doing in the scene)-describe(what) | Predicted: one is washinga chair and the other istaking picturesXGold: yes,on a secondfloor roof deck ,one man iswashing a chair,anotherman is either filming ortaking pictures |
| 4 | does he sit in thechair after washingit？ | Predicted:find (he),find(the chair)→summarize()Gold: find(he),find(the chair)→summarize() | Predicted: where(one man in the video)，where(a chair)→ when(sit in the chairafter washing it)-exist()Gold:where(one man)，where(a chair)→when(sit in the chair after washingit)-exist() | Predicted: no, he does notsit in the chair√Gold: no he does not,thereis a pipe with water runningall over |
| Model | Visual | Count (Loss) | Action (Acc) | Transition (Acc) | FrameQA (Acc) |
| VIS(avg) (Ren et al.,2015a) | R | 4.80 | 0.488 | 0.348 | 0.350 |
| MCB (aggr) (Fukui et al.,2016) | R | 5.17 | 0.589 | 0.243 | 0.257 |
| Yu et al. (Yu et al., 2017) | R | 5.13 | 0.561 | 0.640 | 0.396 |
| ST-VQA (t) (Gao et al., 2018) | R+F | 4.32 | 0.629 | 0.694 | 0.495 |
| Co-Mem (Gao et al.,2018) | R+F | 4.10 | 0.682 | 0.743 | 0.515 |
| PSAC (Li et al.,2019) | R | 4.27 | 0.704 | 0.769 | 0.557 |
| HME (Fan et al.,2019) | R+C | 4.02 | 0.739 | 0.778 | 0.538 |
| STA (Gao et al.,2019) | R | 4.25 | 0.723 | 0.790 | 0.566 |
| CRN+MAC (Le et al.,2019c) | R | 4.23 | 0.713 | 0.787 | 0.592 |
| MQL (Lei et al., 2020) | V | - | - | - | 0.598 |
| QueST (Jiang et al.,2020) | R | 4.19 | 0.759 | 0.810 | 0.597 |
| HGA (Jiang & Han,2020) | R+C | 4.09 | 0.754 | 0.810 | 0.551 |
| GCN (Huang et al., 2020) | R+C | 3.95 | 0.743 | 0.811 | 0.563 |
| HCRN (Le et al.,2020) | R+RX | 3.82 | 0.750 | 0.814 | 0.559 |
| VilNMN | R | 2.65 | 0.845 | 0.887 | 0.747 |
| →soft label programs | R | 1.90 | 0.857 | 0.898 | 0.780 |
| →- res-to-question attn. | R | 3.28 | 0.801 | 0.776 | 0.679 |
| Turn | Question | Dialogue NMN | Video NMN | Response |
| 3 | ...what are theydoing in the scene ？are they on arooftop？ | Predicted:find(they)，find(the scene)→summarize()×Gold: find(they)→summarize() | Predicted: where(two men in thevideo),where(the scene)→when(doing inthe scene)→describe(what)Gold:where(two men)，where (rooftop)→when(doing in the scene)-describe(what) | Predicted: one is washinga chair and the other istaking picturesXGold: yes,on a secondfloor roof deck ,one man iswashing a chair,anotherman is either filming ortaking pictures |
| 4 | does he sit in thechair after washingit？ | Predicted:find (he),find(the chair)→summarize()Gold: find(he),find(the chair)→summarize() | Predicted: where(one man in the video)，where(a chair)→ when(sit in the chairafter washing it)-exist()Gold:where(one man)，where(a chair)→when(sit in the chair after washingit)-exist() | Predicted: no, he does notsit in the chair√Gold: no he does not,thereis a pipe with water runningall over |
| Model | Visual | Count (Loss) | Action (Acc) | Transition (Acc) | FrameQA (Acc) |
| VIS(avg) (Ren et al.,2015a) | R | 4.80 | 0.488 | 0.348 | 0.350 |
| MCB (aggr) (Fukui et al.,2016) | R | 5.17 | 0.589 | 0.243 | 0.257 |
| Yu et al. (Yu et al., 2017) | R | 5.13 | 0.561 | 0.640 | 0.396 |
| ST-VQA (t) (Gao et al., 2018) | R+F | 4.32 | 0.629 | 0.694 | 0.495 |
| Co-Mem (Gao et al.,2018) | R+F | 4.10 | 0.682 | 0.743 | 0.515 |
| PSAC (Li et al.,2019) | R | 4.27 | 0.704 | 0.769 | 0.557 |
| HME (Fan et al.,2019) | R+C | 4.02 | 0.739 | 0.778 | 0.538 |
| STA (Gao et al.,2019) | R | 4.25 | 0.723 | 0.790 | 0.566 |
| CRN+MAC (Le et al.,2019c) | R | 4.23 | 0.713 | 0.787 | 0.592 |
| MQL (Lei et al., 2020) | V | - | - | - | 0.598 |
| QueST (Jiang et al.,2020) | R | 4.19 | 0.759 | 0.810 | 0.597 |
| HGA (Jiang & Han,2020) | R+C | 4.09 | 0.754 | 0.810 | 0.551 |
| GCN (Huang et al., 2020) | R+C | 3.95 | 0.743 | 0.811 | 0.563 |
| HCRN (Le et al.,2020) | R+RX | 3.82 | 0.750 | 0.814 | 0.559 |
| VilNMN | R | 2.65 | 0.845 | 0.887 | 0.747 |
| →soft label programs | R | 1.90 | 0.857 | 0.898 | 0.780 |
| →- res-to-question attn. | R | 3.28 | 0.801 | 0.776 | 0.679 |
| # | Train | Val. | Test | |
| AVSD | Dialogs | 7,659 | 1,787 | 1,710 |
| Turns | 153,180 | 35,740 | 13,490 | |
| Words | 1,450,754 | 339,006 | 110,252 | |
| TGIFQA | Count QA | 24,159 | 2,684 | 3,554 |
| Action QA | 18,428 | 2,047 | 2,274 | |
| Trans. QA | 47,434 | 5,270 | 6,232 | |
| Frame QA | 35,453 | 3,939 | 13,691 |
| # | Train | Val. | Test | |
| AVSD | Dialogs | 7,659 | 1,787 | 1,710 |
| Turns | 153,180 | 35,740 | 13,490 | |
| Words | 1,450,754 | 339,006 | 110,252 | |
| TGIFQA | Count QA | 24,159 | 2,684 | 3,554 |
| Action QA | 18,428 | 2,047 | 2,274 | |
| Trans. QA | 47,434 | 5,270 | 6,232 | |
| Frame QA | 35,453 | 3,939 | 13,691 |
| BLEU4 | CIDEr | |||
| turn position | Model B | Model E | Model B | Model E |
| 1 | 0.579 | 0.587 | 1.623 | 1.650 |
| 2 | 0.429 | 0.430 | 1.155 | 1.142 |
| 3 | 0.275 | 0.289 | 0.867 | 0.846 |
| 4 | 0.309 | 0.305 | 0.859 | 0.855 |
| 5 | 0.355 | 0.335 | 1.088 | 1.023 |
| 6 | 0.357 | 0.329 | 1.044 | 0.950 |
| 7 | 0.342 | 0.325 | 0.896 | 0.847 |
| 8 | 0.361 | 0.332 | 1.025 | 0.973 |
| 9 | 0.383 | 0.431 | 1.043 | 1.182 |
| 10 | 0.395 | 0.371 | 0.931 | 0.977 |
| BLEU4 | CIDEr | |||
| video range (seconds) | Model A | Model B | Model A | Model B |
| 1-23 | 0.432 | 0.447 | 1.298 | 1.355 |
| 23-28 | 0.436 | 0.433 | 1.264 | 1.165 |
| 28-30 | 0.398 | 0.376 | 1.203 | 1.164 |
| 30-30.6 | 0.441 | 0.418 | 1.220 | 1.202 |
| 30.6-31 | 0.413 | 0.411 | 1.250 | 1.166 |
| 31-31.6 | 0.439 | 0.451 | 1.249 | 1.295 |
| 31.6-32 | 0.430 | 0.419 | 1.217 | 1.192 |
| 32-33 | 0.468 | 0.445 | 1.343 | 1.237 |
| 33-37 | 0.388 | 0.381 | 1.149 | 1.124 |
| 37-75 | 0.356 | 0.365 | 0.910 | 0.962 |
| BLEU4 | CIDEr | |||
| turn position | Model B | Model E | Model B | Model E |
| 1 | 0.579 | 0.587 | 1.623 | 1.650 |
| 2 | 0.429 | 0.430 | 1.155 | 1.142 |
| 3 | 0.275 | 0.289 | 0.867 | 0.846 |
| 4 | 0.309 | 0.305 | 0.859 | 0.855 |
| 5 | 0.355 | 0.335 | 1.088 | 1.023 |
| 6 | 0.357 | 0.329 | 1.044 | 0.950 |
| 7 | 0.342 | 0.325 | 0.896 | 0.847 |
| 8 | 0.361 | 0.332 | 1.025 | 0.973 |
| 9 | 0.383 | 0.431 | 1.043 | 1.182 |
| 10 | 0.395 | 0.371 | 0.931 | 0.977 |
| BLEU4 | CIDEr | |||
| video range (seconds) | Model A | Model B | Model A | Model B |
| 1-23 | 0.432 | 0.447 | 1.298 | 1.355 |
| 23-28 | 0.436 | 0.433 | 1.264 | 1.165 |
| 28-30 | 0.398 | 0.376 | 1.203 | 1.164 |
| 30-30.6 | 0.441 | 0.418 | 1.220 | 1.202 |
| 30.6-31 | 0.413 | 0.411 | 1.250 | 1.166 |
| 31-31.6 | 0.439 | 0.451 | 1.249 | 1.295 |
| 31.6-32 | 0.430 | 0.419 | 1.217 | 1.192 |
| 32-33 | 0.468 | 0.445 | 1.343 | 1.237 |
| 33-37 | 0.388 | 0.381 | 1.149 | 1.124 |
| 37-75 | 0.356 | 0.365 | 0.910 | 0.962 |
| Turn | Question | Dialogue UnderstandingProgram | Video UnderstandingProgram | Response |
| 1 | hi how many peopleare in the video fromstart to finish ? | Predicted: summari ze()Gold: summarize () | Predicted: describe (how manypeople)Gold:describe (how manypeople) | Predicted: there are twopeople in the video.Gold: hello,there aretwo men |
| 2 | are they carrying ona conversationbetween themselves？ | Predicted:find(they)→summarize()Gold:find(they)，find(themselves)→summarize() | Predicted: where(two men in the←video)→when(carrying on aconversation betweenthemselves)-exist()Gold:where(two men)→ when(carrying a conversationbetween themselves)-exist() | Predicted: yes,they arecarrying a conversation×Gold: no,backgroundnoise is birdsong andshuffling noises |
| 3 | what are they doingin the scene ?arethey on a rooftop ? | Predicted: find(they)，find(the scene)→summarize() ×Gold: find(they)→summarize() | Predicted:where(two men in thevideo),where(the scene)→when(doing in the scene)→describe(what) ×Gold: where(two men),where(rooftop)→when(doingin the scene)-describe(what) | Predicted: one iswashing a chair and theother is taking pictures×Gold: yes,on a secondfloor roof deck ,oneman is washing a chair,another man is eitherfilming or taking pictures |
| 4 | does he sit in thechair after washingit？ | Predicted: find(he),find(the chair)→summarize()Gold: find (he),find(the chair)→summarize() | Predicted: where(one man in thevideo)，where(a chair)→when(sit in the chair afterwashing it)→exist()Gold:where(one man)，where(achair)→ when(sit in thechair after washing it)→exist() | Predicted: no,he doesnot sit in the chairGold: no he does not,there is a pipe withwater running all over |
| Turn | Question | Dialogue UnderstandingProgram | Video UnderstandingProgram | Response |
| 1 | hi how many peopleare in the video fromstart to finish ? | Predicted: summari ze()Gold: summarize () | Predicted: describe (how manypeople)Gold:describe (how manypeople) | Predicted: there are twopeople in the video.Gold: hello,there aretwo men |
| 2 | are they carrying ona conversationbetween themselves？ | Predicted:find(they)→summarize()Gold:find(they)，find(themselves)→summarize() | Predicted: where(two men in the←video)→when(carrying on aconversation betweenthemselves)-exist()Gold:where(two men)→ when(carrying a conversationbetween themselves)-exist() | Predicted: yes,they arecarrying a conversation×Gold: no,backgroundnoise is birdsong andshuffling noises |
| 3 | what are they doingin the scene ?arethey on a rooftop ? | Predicted: find(they)，find(the scene)→summarize() ×Gold: find(they)→summarize() | Predicted:where(two men in thevideo),where(the scene)→when(doing in the scene)→describe(what) ×Gold: where(two men),where(rooftop)→when(doingin the scene)-describe(what) | Predicted: one iswashing a chair and theother is taking pictures×Gold: yes,on a secondfloor roof deck ,oneman is washing a chair,another man is eitherfilming or taking pictures |
| 4 | does he sit in thechair after washingit？ | Predicted: find(he),find(the chair)→summarize()Gold: find (he),find(the chair)→summarize() | Predicted: where(one man in thevideo)，where(a chair)→when(sit in the chair afterwashing it)→exist()Gold:where(one man)，where(achair)→ when(sit in thechair after washing it)→exist() | Predicted: no,he doesnot sit in the chairGold: no he does not,there is a pipe withwater running all over |
| CSL | 4 CYCLES | 6 CYCLES | CONNECTIVITY | Multi-task | Invariant to vertex permutations? | ||||
| (Accuracy %%)↑ | (Accuracy % %) ↑ | (Accuracy %)↑ | 8 CYCLES (Accuracy %)↑ | (logMSE)↓ | DIAMETER (logMSE)↓ | SPECTRAL RADIUS (logMSE)↓ | |||
| (n-1) | GIN (orig.) | 4.66 ± 4.00 | 93.0t | 92.7t | 92.5t | -3.419 ± 0.320 | 0.588 ± 0.354 | -2.130 ± 1.396 | |
| 88.66 ± 22.66 | 95.17 ± 4.91 | 97.35 ± 0.74 | 94.69 ± 2.34 | -3.575 ± 0.395 | -0.195 ± 0.714 | -2.732 ± 0.793 | |||
| Reeooor | (n-2) | 78.66 ± 22.17 | 94.06±5.10 | 97.50±0.72 | 95.04 ± 2.69 | -3.799 ± 0.187 | -0.207 ± 0.381 | -2.344 ± 0.569 | |
| (n-3) | 73.33 ± 16.19 | 96.61 ± 1.40 | 97.84 ± 1.37 | 94.48 ± 2.13 | -3.779 ± 0.064 | 0.105 ±0.225 | -1.908 ± 0.860 | <<<< | |
| [n/2] | 40.66 ± 9.04 | 75.13 ± 0.26 | 63.28 ± 0.59 | 63.53 ± 1.14 | -3.765 ± 0.083 | 0.564 ± 0.025 | -2.130 ± 0.166 | ||
| GCN(orig.) | 6.66 ± 2.10 | 98.336 ± 0.24 | 95.73 ± 2.72 | 87.14 ± 12.73 | -3.781 ± 0.075 | 0.087 ± 0.186 | -2.204 ± 0.362 | ||
| (n-1) | 100.00 ±0.00 | 99.00 ±0.10 | 97.63 ± 0.19 | 94.99 ± 2.31 | -4.039 ± 0.101 | -1.175 ± 0.425 | -3.625 ± 0.536 | ||
| (n-2) | 100.00±0.00 | 98.77 ±0.61 | 97.89 ± 0.69 | 97.82 ± 1.10 | -3.970 ± 0.059 | -0.577 ± 0.135 | -3.397 ± 0.273 | ||
| Reeeoss | (n-3) | 96.00 ±6.46 | 99.11 ± 0.19 | 98.31 ± 0.52 | 97.18 ± 0.58 | -3.995 ± 0.031 | -0.333 ± 0.117 | -3.105 ± 0.286 | |
| [n/2] | 49.33 ± 7.42 | 75.19 ± 0.19 | 66.04 ± 0.59 | 63.66 ± 0.51 | -3.693 ± 0.063 | 0.8518 ± 0.016 | -1.838 ± 0.054 | ||
| PNA (orig.) | 10.00 ± 2.98 | 81.59 ± 19.86 | 95.57 ± 0.36 | 84.81 ± 16.48 | -3.794 ± 0.155 | -0.605 ± 0.097 | -3.610 ± 0.137 | ||
| (n-1) | 100.00 ±0.00 | 97.88 ± 2.19 | 99.18 ±0.20 | 98.92 ± 0.72 | -3.904 ± 0.001 | -0.765 ± 0.032 | -3.954 ±0.118 | ||
| Rreeoor (n-2) | 95.33 ± 7.77 | 99.12 ± 0.28 | 99.10 ± 0.57 | 99.22 ± 0.27 | -3.781 ± 0.085 | -0.090 ± 0.135 | -3.478 ± 0.206 | ||
| (n-3) | 95.33 ± 5.81 | 89.36 ± 0.22 75.34 ±0.18 | 99.34 ± 0.26 | 93.92 ± 8.15 64.01 ± 0.30 | -3.710 ± 0.209 -2.977 ± 0.065 | 0.042 ± 0.047 | -3.311 ± 0.067 | √ | |
| [n/2] | 42.66 ± 11.03 | 65.58 ± 0.95 | 1.445 ± 0.037 | -1.073 ± 0.075 | |||||
| Positional GIN Ring-GNN | 99.33+ ± 1.33 | 88.3† | 96.1† | 95.3t | -1.61t | -2.17 | -2.66t | x<ν | |
| 10.00 ± 0.00 | 99.9† | 100.0† | 71.4 | 二 | |||||
| PPGN (3-WL) | 97.80 ± 10.91 | 99.8† | 87.1† | 76.5† | 二 | 二 |
| OGBG-MOLTOX21 OGBG-MOLTOXCAST OGBG-MOLFREESOLV OGBG-MOLESOL 0GBG-MOLLIPO OGBG-MOLPCBA(ROC-AUC %)↑ (ROC-AUC %) ↑ (RSMSE)↓ (RSMSE)↓ (RSMSE)↓ (AP%)↑ | |||||||||||
| GIN (orig.) | 74.91 ± 0.51 | 63.41 ± 0.74 | 2.411 ± 0.123 1.111 ± 0.038 0.754 ± 0.010 21.16 ± 0.28 | ||||||||
| (n-1) | 75.15 ± 1.40 | 63.95 ± 0.53 | 2.283 ± 0.279 | 1.026 ± 0.033 | 0.716± 0.020 | 23.60 ± 0.02 | |||||
| (n-2) | 76.84 ± 0.62 | 65.36 ± 0.49 | 2.117 ± 0.181 | 1.006 ± 0.030 | 0.736± 0.025 | 23.25 ± 0.00 | |||||
| (n-3) | 76.78 ± 0.64 | 64.84 ± 0.71 | 2.370 ± 0.326 | 1.055 ± 0.031 | 0.738 ± 0.018 | 23.33 ± 0.09 | |||||
| [n/2] | 74.40 ± 0.75 | 62.29 ± 0.28 | 2.531 ± 0.206 | 1.343 ± 0.053 | 0.842 ± 0.020 | 13.50 ± 0.32 | |||||
| GCN (orig.) | 75.29 ± 0.69 | 63.54 ± 0.42 | 2.417 ± 0.178 | 1.106 ± 0.036 | 0.793 ± 0.040 | 20.20 ± 0.24 | |||||
| 上 | (n-1) | 76.46 ± 0.77 | 64.51 ± 0.60 | 2.524 ± 0.300 | 1.096 ± 0.045 | 0.760 ± 0.015 | 21.25 ± 0.25 | ||||
| (n-2) | 75.58 ± 0.99 | 64.38 ± 0.39 | 2.467 ± 0.231 | 1.086 ± 0.048 | 0.766 ± 0.025 | 20.10 ± 0.08 | |||||
| (n-3) | 75.88 ± 0.73 | 64.70 ±0.81 | 2.345 ± 0.261 | 1.114 ± 0.047 | 0.754 ± 0.021 | 19.04 ± 0.03 | |||||
| [n/2] | 74.03 ± 0.63 | 62.80 ± 0.77 | 2.599 ± 0.161 | 1.372 ± 0.048 | 0.835± 0.020 | 11.69 ± 1.41 | |||||
| PNA (orig.) | 74.28 ± 0.52 | 62.69 ± 0.63 | 2.192 ± 0.125 | 1.140 ± 0.032 | 0.759 ± 0.017 | 25.45 ± 0.04 | |||||
| (n-1) | 73.64 ± 0.74 | 64.14 ±0.76 | 2.341 ± 0.070 | 1.723 ± 0.145 | 0.743 ± 0.015 | 23.11 ± 0.05 | |||||
| (n-2) | 74.89 ± 0.29 | 65.22 ± 0.47 | 2.298 ± 0.115 | 1.392 ± 0.272 | 0.794 ± 0.065 | 22.10±0.03 | |||||
| (n-3) | 75.10 ± 0.73 | 65.03 ± 0.58 | 2.133 ± 0.086 | 1.360 ± 0.163 | 0.785 ± 0.041 | 20.05 ± 0.15 | |||||
| [n/2] | 73.71 ± 0.61 | 61.25 ± 0.49 | 2.185 ± 0.231 | 1.157 ± 0.056 | 0.843 ± 0.018 | 12.33 ± 1.20 | |||||
| CSL | 4 CYCLES | 6 CYCLES | CONNECTIVITY | Multi-task | Invariant to vertex permutations? | ||||
| (Accuracy %%)↑ | (Accuracy % %) ↑ | (Accuracy %)↑ | 8 CYCLES (Accuracy %)↑ | (logMSE)↓ | DIAMETER (logMSE)↓ | SPECTRAL RADIUS (logMSE)↓ | |||
| (n-1) | GIN (orig.) | 4.66 ± 4.00 | 93.0t | 92.7t | 92.5t | -3.419 ± 0.320 | 0.588 ± 0.354 | -2.130 ± 1.396 | |
| 88.66 ± 22.66 | 95.17 ± 4.91 | 97.35 ± 0.74 | 94.69 ± 2.34 | -3.575 ± 0.395 | -0.195 ± 0.714 | -2.732 ± 0.793 | |||
| Reeooor | (n-2) | 78.66 ± 22.17 | 94.06±5.10 | 97.50±0.72 | 95.04 ± 2.69 | -3.799 ± 0.187 | -0.207 ± 0.381 | -2.344 ± 0.569 | |
| (n-3) | 73.33 ± 16.19 | 96.61 ± 1.40 | 97.84 ± 1.37 | 94.48 ± 2.13 | -3.779 ± 0.064 | 0.105 ±0.225 | -1.908 ± 0.860 | <<<< | |
| [n/2] | 40.66 ± 9.04 | 75.13 ± 0.26 | 63.28 ± 0.59 | 63.53 ± 1.14 | -3.765 ± 0.083 | 0.564 ± 0.025 | -2.130 ± 0.166 | ||
| GCN(orig.) | 6.66 ± 2.10 | 98.336 ± 0.24 | 95.73 ± 2.72 | 87.14 ± 12.73 | -3.781 ± 0.075 | 0.087 ± 0.186 | -2.204 ± 0.362 | ||
| (n-1) | 100.00 ±0.00 | 99.00 ±0.10 | 97.63 ± 0.19 | 94.99 ± 2.31 | -4.039 ± 0.101 | -1.175 ± 0.425 | -3.625 ± 0.536 | ||
| (n-2) | 100.00±0.00 | 98.77 ±0.61 | 97.89 ± 0.69 | 97.82 ± 1.10 | -3.970 ± 0.059 | -0.577 ± 0.135 | -3.397 ± 0.273 | ||
| Reeeoss | (n-3) | 96.00 ±6.46 | 99.11 ± 0.19 | 98.31 ± 0.52 | 97.18 ± 0.58 | -3.995 ± 0.031 | -0.333 ± 0.117 | -3.105 ± 0.286 | |
| [n/2] | 49.33 ± 7.42 | 75.19 ± 0.19 | 66.04 ± 0.59 | 63.66 ± 0.51 | -3.693 ± 0.063 | 0.8518 ± 0.016 | -1.838 ± 0.054 | ||
| PNA (orig.) | 10.00 ± 2.98 | 81.59 ± 19.86 | 95.57 ± 0.36 | 84.81 ± 16.48 | -3.794 ± 0.155 | -0.605 ± 0.097 | -3.610 ± 0.137 | ||
| (n-1) | 100.00 ±0.00 | 97.88 ± 2.19 | 99.18 ±0.20 | 98.92 ± 0.72 | -3.904 ± 0.001 | -0.765 ± 0.032 | -3.954 ±0.118 | ||
| Rreeoor (n-2) | 95.33 ± 7.77 | 99.12 ± 0.28 | 99.10 ± 0.57 | 99.22 ± 0.27 | -3.781 ± 0.085 | -0.090 ± 0.135 | -3.478 ± 0.206 | ||
| (n-3) | 95.33 ± 5.81 | 89.36 ± 0.22 75.34 ±0.18 | 99.34 ± 0.26 | 93.92 ± 8.15 64.01 ± 0.30 | -3.710 ± 0.209 -2.977 ± 0.065 | 0.042 ± 0.047 | -3.311 ± 0.067 | √ | |
| [n/2] | 42.66 ± 11.03 | 65.58 ± 0.95 | 1.445 ± 0.037 | -1.073 ± 0.075 | |||||
| Positional GIN Ring-GNN | 99.33+ ± 1.33 | 88.3† | 96.1† | 95.3t | -1.61t | -2.17 | -2.66t | x<ν | |
| 10.00 ± 0.00 | 99.9† | 100.0† | 71.4 | 二 | |||||
| PPGN (3-WL) | 97.80 ± 10.91 | 99.8† | 87.1† | 76.5† | 二 | 二 |
| OGBG-MOLTOX21 OGBG-MOLTOXCAST OGBG-MOLFREESOLV OGBG-MOLESOL 0GBG-MOLLIPO OGBG-MOLPCBA(ROC-AUC %)↑ (ROC-AUC %) ↑ (RSMSE)↓ (RSMSE)↓ (RSMSE)↓ (AP%)↑ | |||||||||||
| GIN (orig.) | 74.91 ± 0.51 | 63.41 ± 0.74 | 2.411 ± 0.123 1.111 ± 0.038 0.754 ± 0.010 21.16 ± 0.28 | ||||||||
| (n-1) | 75.15 ± 1.40 | 63.95 ± 0.53 | 2.283 ± 0.279 | 1.026 ± 0.033 | 0.716± 0.020 | 23.60 ± 0.02 | |||||
| (n-2) | 76.84 ± 0.62 | 65.36 ± 0.49 | 2.117 ± 0.181 | 1.006 ± 0.030 | 0.736± 0.025 | 23.25 ± 0.00 | |||||
| (n-3) | 76.78 ± 0.64 | 64.84 ± 0.71 | 2.370 ± 0.326 | 1.055 ± 0.031 | 0.738 ± 0.018 | 23.33 ± 0.09 | |||||
| [n/2] | 74.40 ± 0.75 | 62.29 ± 0.28 | 2.531 ± 0.206 | 1.343 ± 0.053 | 0.842 ± 0.020 | 13.50 ± 0.32 | |||||
| GCN (orig.) | 75.29 ± 0.69 | 63.54 ± 0.42 | 2.417 ± 0.178 | 1.106 ± 0.036 | 0.793 ± 0.040 | 20.20 ± 0.24 | |||||
| 上 | (n-1) | 76.46 ± 0.77 | 64.51 ± 0.60 | 2.524 ± 0.300 | 1.096 ± 0.045 | 0.760 ± 0.015 | 21.25 ± 0.25 | ||||
| (n-2) | 75.58 ± 0.99 | 64.38 ± 0.39 | 2.467 ± 0.231 | 1.086 ± 0.048 | 0.766 ± 0.025 | 20.10 ± 0.08 | |||||
| (n-3) | 75.88 ± 0.73 | 64.70 ±0.81 | 2.345 ± 0.261 | 1.114 ± 0.047 | 0.754 ± 0.021 | 19.04 ± 0.03 | |||||
| [n/2] | 74.03 ± 0.63 | 62.80 ± 0.77 | 2.599 ± 0.161 | 1.372 ± 0.048 | 0.835± 0.020 | 11.69 ± 1.41 | |||||
| PNA (orig.) | 74.28 ± 0.52 | 62.69 ± 0.63 | 2.192 ± 0.125 | 1.140 ± 0.032 | 0.759 ± 0.017 | 25.45 ± 0.04 | |||||
| (n-1) | 73.64 ± 0.74 | 64.14 ±0.76 | 2.341 ± 0.070 | 1.723 ± 0.145 | 0.743 ± 0.015 | 23.11 ± 0.05 | |||||
| (n-2) | 74.89 ± 0.29 | 65.22 ± 0.47 | 2.298 ± 0.115 | 1.392 ± 0.272 | 0.794 ± 0.065 | 22.10±0.03 | |||||
| (n-3) | 75.10 ± 0.73 | 65.03 ± 0.58 | 2.133 ± 0.086 | 1.360 ± 0.163 | 0.785 ± 0.041 | 20.05 ± 0.15 | |||||
| [n/2] | 73.71 ± 0.61 | 61.25 ± 0.49 | 2.185 ± 0.231 | 1.157 ± 0.056 | 0.843 ± 0.018 | 12.33 ± 1.20 | |||||
| OGBG-MOLTOX21 OGBG-MOLTOXCAST OGBG-MOLFREESOLV OGBG-MOLESOL 0GBG-MOLLIPO OGBG-MOLPCBA(ROC-AUC %)↑ (ROC-AUC %) ↑ (RSMSE)↓ (RSMSE)↓ (RSMSE)↓ (AP%)↑ | |||||||||||
| GIN (orig.) | 74.91 ± 0.51 | 63.41 ± 0.74 | 2.411 ± 0.123 1.111 ± 0.038 0.754 ± 0.010 21.16 ± 0.28 | ||||||||
| (n-1) | 75.15 ± 1.40 | 63.95 ± 0.53 | 2.283 ± 0.279 | 1.026 ± 0.033 | 0.716± 0.020 | 23.60 ± 0.02 | |||||
| (n-2) | 76.84 ± 0.62 | 65.36 ± 0.49 | 2.117 ± 0.181 | 1.006 ± 0.030 | 0.736± 0.025 | 23.25 ± 0.00 | |||||
| (n-3) | 76.78 ± 0.64 | 64.84 ± 0.71 | 2.370 ± 0.326 | 1.055 ± 0.031 | 0.738 ± 0.018 | 23.33 ± 0.09 | |||||
| [n/2] | 74.40 ± 0.75 | 62.29 ± 0.28 | 2.531 ± 0.206 | 1.343 ± 0.053 | 0.842 ± 0.020 | 13.50 ± 0.32 | |||||
| GCN (orig.) | 75.29 ± 0.69 | 63.54 ± 0.42 | 2.417 ± 0.178 | 1.106 ± 0.036 | 0.793 ± 0.040 | 20.20 ± 0.24 | |||||
| 上 | (n-1) | 76.46 ± 0.77 | 64.51 ± 0.60 | 2.524 ± 0.300 | 1.096 ± 0.045 | 0.760 ± 0.015 | 21.25 ± 0.25 | ||||
| (n-2) | 75.58 ± 0.99 | 64.38 ± 0.39 | 2.467 ± 0.231 | 1.086 ± 0.048 | 0.766 ± 0.025 | 20.10 ± 0.08 | |||||
| (n-3) | 75.88 ± 0.73 | 64.70 ±0.81 | 2.345 ± 0.261 | 1.114 ± 0.047 | 0.754 ± 0.021 | 19.04 ± 0.03 | |||||
| [n/2] | 74.03 ± 0.63 | 62.80 ± 0.77 | 2.599 ± 0.161 | 1.372 ± 0.048 | 0.835± 0.020 | 11.69 ± 1.41 | |||||
| PNA (orig.) | 74.28 ± 0.52 | 62.69 ± 0.63 | 2.192 ± 0.125 | 1.140 ± 0.032 | 0.759 ± 0.017 | 25.45 ± 0.04 | |||||
| (n-1) | 73.64 ± 0.74 | 64.14 ±0.76 | 2.341 ± 0.070 | 1.723 ± 0.145 | 0.743 ± 0.015 | 23.11 ± 0.05 | |||||
| (n-2) | 74.89 ± 0.29 | 65.22 ± 0.47 | 2.298 ± 0.115 | 1.392 ± 0.272 | 0.794 ± 0.065 | 22.10±0.03 | |||||
| (n-3) | 75.10 ± 0.73 | 65.03 ± 0.58 | 2.133 ± 0.086 | 1.360 ± 0.163 | 0.785 ± 0.041 | 20.05 ± 0.15 | |||||
| [n/2] | 73.71 ± 0.61 | 61.25 ± 0.49 | 2.185 ± 0.231 | 1.157 ± 0.056 | 0.843 ± 0.018 | 12.33 ± 1.20 | |||||
| CSL | 4 CYCLES | 6 CYCLES | CONNECTIVITY | Multi-task | Invariant to vertex permutations? | ||||
| (Accuracy %%)↑ | (Accuracy % %) ↑ | (Accuracy %)↑ | 8 CYCLES (Accuracy %)↑ | (logMSE)↓ | DIAMETER (logMSE)↓ | SPECTRAL RADIUS (logMSE)↓ | |||
| (n-1) | GIN (orig.) | 4.66 ± 4.00 | 93.0t | 92.7t | 92.5t | -3.419 ± 0.320 | 0.588 ± 0.354 | -2.130 ± 1.396 | |
| 88.66 ± 22.66 | 95.17 ± 4.91 | 97.35 ± 0.74 | 94.69 ± 2.34 | -3.575 ± 0.395 | -0.195 ± 0.714 | -2.732 ± 0.793 | |||
| Reeooor | (n-2) | 78.66 ± 22.17 | 94.06±5.10 | 97.50±0.72 | 95.04 ± 2.69 | -3.799 ± 0.187 | -0.207 ± 0.381 | -2.344 ± 0.569 | |
| (n-3) | 73.33 ± 16.19 | 96.61 ± 1.40 | 97.84 ± 1.37 | 94.48 ± 2.13 | -3.779 ± 0.064 | 0.105 ±0.225 | -1.908 ± 0.860 | <<<< | |
| [n/2] | 40.66 ± 9.04 | 75.13 ± 0.26 | 63.28 ± 0.59 | 63.53 ± 1.14 | -3.765 ± 0.083 | 0.564 ± 0.025 | -2.130 ± 0.166 | ||
| GCN(orig.) | 6.66 ± 2.10 | 98.336 ± 0.24 | 95.73 ± 2.72 | 87.14 ± 12.73 | -3.781 ± 0.075 | 0.087 ± 0.186 | -2.204 ± 0.362 | ||
| (n-1) | 100.00 ±0.00 | 99.00 ±0.10 | 97.63 ± 0.19 | 94.99 ± 2.31 | -4.039 ± 0.101 | -1.175 ± 0.425 | -3.625 ± 0.536 | ||
| (n-2) | 100.00±0.00 | 98.77 ±0.61 | 97.89 ± 0.69 | 97.82 ± 1.10 | -3.970 ± 0.059 | -0.577 ± 0.135 | -3.397 ± 0.273 | ||
| Reeeoss | (n-3) | 96.00 ±6.46 | 99.11 ± 0.19 | 98.31 ± 0.52 | 97.18 ± 0.58 | -3.995 ± 0.031 | -0.333 ± 0.117 | -3.105 ± 0.286 | |
| [n/2] | 49.33 ± 7.42 | 75.19 ± 0.19 | 66.04 ± 0.59 | 63.66 ± 0.51 | -3.693 ± 0.063 | 0.8518 ± 0.016 | -1.838 ± 0.054 | ||
| PNA (orig.) | 10.00 ± 2.98 | 81.59 ± 19.86 | 95.57 ± 0.36 | 84.81 ± 16.48 | -3.794 ± 0.155 | -0.605 ± 0.097 | -3.610 ± 0.137 | ||
| (n-1) | 100.00 ±0.00 | 97.88 ± 2.19 | 99.18 ±0.20 | 98.92 ± 0.72 | -3.904 ± 0.001 | -0.765 ± 0.032 | -3.954 ±0.118 | ||
| Rreeoor (n-2) | 95.33 ± 7.77 | 99.12 ± 0.28 | 99.10 ± 0.57 | 99.22 ± 0.27 | -3.781 ± 0.085 | -0.090 ± 0.135 | -3.478 ± 0.206 | ||
| (n-3) | 95.33 ± 5.81 | 89.36 ± 0.22 75.34 ±0.18 | 99.34 ± 0.26 | 93.92 ± 8.15 64.01 ± 0.30 | -3.710 ± 0.209 -2.977 ± 0.065 | 0.042 ± 0.047 | -3.311 ± 0.067 | √ | |
| [n/2] | 42.66 ± 11.03 | 65.58 ± 0.95 | 1.445 ± 0.037 | -1.073 ± 0.075 | |||||
| Positional GIN Ring-GNN | 99.33+ ± 1.33 | 88.3† | 96.1† | 95.3t | -1.61t | -2.17 | -2.66t | x<ν | |
| 10.00 ± 0.00 | 99.9† | 100.0† | 71.4 | 二 | |||||
| PPGN (3-WL) | 97.80 ± 10.91 | 99.8† | 87.1† | 76.5† | 二 | 二 |
| Strategy | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 |
| No Attack | 29.3 | 33.4 | 38.1 | 40.7 | 42.1 | 42.5 | 45.7 | 46.9 | 53.9 |
| Size | 26.0 | 14.7 | 31.9 | 32.5 | 35.6 | 35.4 | 38.6 | 40.0 | 47.8 |
| Localization | 22.8 | 6.4 | 27.4 | 28.1 | 31.7 | 30.8 | 34.4 | 35.9 | 45.1 |
| Classification | 18.1 | 1.2 | 19.9 | 20.5 | 24.3 | 22.6 | 26.4 | 28.2 | 39.9 |
| Technique | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 |
| None | 18.1 | 1.2 | 19.9 | 20.5 | 24.3 | 22.6 | 26.4 | 28.2 | 39.9 |
| DI | 18.1 | 1.0 | 19.9 | 20.5 | 23.9 | 22.4 | 26.3 | 27.9 | 39.6 |
| TI | 17.0 | 2.4 | 20.8 | 20.8 | 25.2 | 23.0 | 27.9 | 29.7 | 41.5 |
| SI | 14.6 | 0.7 | 16.3 | 17.0 | 20.4 | 19.1 | 22.3 | 23.8 | 35.0 |
| Method | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | |
| Reference | No Attack Ablation | 29.3 24.9 | 33.4 31.4 | 38.1 31.2 | 40.7 31.6 | 42.1 35.0 | 42.5 34.3 | 45.7 37.5 | 46.9 38.8 | 53.9 48.6 |
| Classification Attack | PGD SI-PGD | 26.4 27.5 | 30.4 31.6 | 34.4 36.1 | 35.4 37.1 | 38.4 | 38.3 | 41.7 | 43.1 | 51.1 |
| MI-DI-PGD | 22.9 | 26.2 | 40.0 | 40.1 | 43.5 | 44.8 | 52.4 | |||
| MI-TI-PGD | 29.3 | 30.0 | 33.2 | 32.1 | 36.0 | 37.5 | 48.0 | |||
| CD-painting | 20.1 16.4 | 23.7 20.8 | 24.9 21.3 | 25.4 22.8 | 30.1 26.6 | 27.4 24.5 | 32.8 28.9 | 34.5 29.5 | 47.1 42.3 | |
| Detection Attack | CD-comics | 16.6 | 21.6 | 21.7 | 22.7 | 26.8 | 24.3 | 29.1 | 42.3 | 43.7 |
| Dfool | 23.3 | 2.5 | 29.2 | 29.8 | 33.3 | 32.9 | 36.5 | 38.0 | 47.5 | |
| Loc DAG | 21.9 20.8 | 0.2 | 25.8 | 26.6 | 29.8 | 29.4 | 33.2 | 33.2 | 45.2 | |
| Ours | RAD | 14.6 | 0.6 0.7 | 22.8 16.3 | 23.4 17.0 | 26.8 20.4 | 25.6 19.1 | 28.9 22.3 | 31.0 23.8 | 40.6 35.0 |
| mAP | mAP50 | mAP75 | |||||||
| Surrogate | M5 | M7 | M8 | M5 | M7 | M8 | M5 | M7 | M8 |
| None | 38.0 | 39.4 | 40.8 | 60.6 | 61.3 | 63.3 | 40.9 | 42.9 | 44.1 |
| Ablation | 31.0 | 31.9 | 33.5 | 51.2 | 51.0 | 53.7 | 32.4 | 34.3 | 35.4 |
| M2 | 17.9 | 18.6 | 20.3 | 31.6 | 31.7 | 34.5 | 18.0 | 18.9 | 20.7 |
| M3 | 11.6 | 11.9 | 12.9 | 19.2 | 19.1 | 20.7 | 12.1 | 12.6 | 13.7 |
| M5 | 1.2 | 11.1 | 11.8 | 2.4 | 17.9 | 18.9 | 1.0 | 11.9 | 12.6 |
| RMSE | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | |
| COCO detection AOCO detection | 0.000 6.469 | 29.3 14.6 | 33.4 0.7 | 38.1 16.3 | 40.7 17 | 42.1 20.4 | 42.5 19.1 | 45.7 22.3 | 46.9 23.8 | 53.9 35.0 |
| COCO segmentation | 0.000 | N | 38.0 | 39.4 | 40.8 | \\ | ||||
| AOCO segmentation | 6.606 | 广 | 广 | \\ | 1.2 | 一 | 11.1 | 11.8 |
| ID | Model | Backbone | mAP |
| M1 | SSD512 (Liu et al. (2016)) | VGG16 | 29.3 |
| M2 | YOLOv3 (Redmon & Farhadi (2018)) | Darknet | 33.4 |
| M3 | RetinaNet (Lin et al. (2017)) | ResNet-101 | 38.1 |
| M4 | Faster R-CNN (Ren et al. (2015)) | ResNeXt-101-64*4d | 40.7 |
| M5 | Mask R-CNN (He et al. (2017)) | ResNeXt-101-64*4d | 42.1 |
| M6 | Cascade RCNN (Cai & Vasconcelos (2018)) | ResNet-101 | 42.5 |
| M7 | Cascade Mask R-CNN (Cai & Vasconcelos (2018)) | ResNeXt-101-64*4d | 45.7 |
| M8 | Hyrbrid Task Cascade (Chen et al. (2019a)) | ResNeXt-101-64*4d | 46.9 |
| M9 | EfficientDet (Tan et al. (2020)) | EfficientNet+BiFPN | 53.9 |
| Strategy | Parameter | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 |
| Dynamic | -1 | 18.3 | 1.2 | 20.0 | 20.6 | 24.3 | 22.8 | 26.7 | 28.3 | 40.0 |
| -2 | 18.2 | 1.2 | 20.1 | 20.7 | 24.2 | 22.8 | 26.6 | 28.0 | 39.8 | |
| -3 | 18.4 | 1.3 | 20.3 | 20.8 | 24.3 | 22.9 | 26.7 | 28.5 | 40.2 | |
| Static | 10 | 18.2 | 1.1 | 19.9 | 20.5 | 24.2 | 22.8 | 26.2 | 28.0 | 39.8 |
| 20 | 18.1 | 1.2 | 19.9 | 20.5 | 24.3 | 22.6 | 26.4 | 28.2 | 39.9 | |
| 30 | 18.2 | 1.3 | 20.1 | 20.7 | 24.3 | 22.8 | 26.0 | 27.9 | 40.1 |
| Surrogate | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 |
| YOLOv3 | 14.6 | 0.7 | 16.3 | 17.0 | 20.4 | 19.1 | 22.3 | 23.8 | 35.0 |
| Mask R-CNN | 20.4 | 24.2 | 25.7 | 26.5 | 1.1 | 28.9 | 33.1 | 34.9 | 45.4 |
| RetinaNet | 20.7 | 6.1 | 2.3 | 25.7 | 29.3 | 28.2 | 31.7 | 33.8 | 44.1 |
| M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | |
| COCO detection AOCO detection | 49.2 26.7 | 56.4 1.6 | 58.1 27.6 | 62.0 29.1 | 63.8 34.5 | 60.7 29.9 | 64.1 34.3 | 66.0 37.4 | 74.3 51.7 |
| COCO segmentation | 人 | 一 | 60.6 | 一 | 61.3 | 63.3 | 一 | ||
| AOCO segmentation | 2.4 | 一 | 17.9 | 18.9 | 一 |
| M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | |
| COCO detection AOCO detection | 30.8 14.2 | 35.8 0.6 | 40.6 16.5 | 44.6 17.1 | 46.3 20.8 | 46.3 19.9 | 50.0 23.3 | 51.2 24.6 | 59.9 37.4 |
| COCO segmentation | / | / | 40.9 | 一 | 42.9 | 44.1 | 一 | ||
| AOCO segmentation | 1.0 | / | 11.9 | 12.6 |
| Strategy | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 |
| No Attack | 29.3 | 33.4 | 38.1 | 40.7 | 42.1 | 42.5 | 45.7 | 46.9 | 53.9 |
| Size | 26.0 | 14.7 | 31.9 | 32.5 | 35.6 | 35.4 | 38.6 | 40.0 | 47.8 |
| Localization | 22.8 | 6.4 | 27.4 | 28.1 | 31.7 | 30.8 | 34.4 | 35.9 | 45.1 |
| Classification | 18.1 | 1.2 | 19.9 | 20.5 | 24.3 | 22.6 | 26.4 | 28.2 | 39.9 |
| Technique | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 |
| None | 18.1 | 1.2 | 19.9 | 20.5 | 24.3 | 22.6 | 26.4 | 28.2 | 39.9 |
| DI | 18.1 | 1.0 | 19.9 | 20.5 | 23.9 | 22.4 | 26.3 | 27.9 | 39.6 |
| TI | 17.0 | 2.4 | 20.8 | 20.8 | 25.2 | 23.0 | 27.9 | 29.7 | 41.5 |
| SI | 14.6 | 0.7 | 16.3 | 17.0 | 20.4 | 19.1 | 22.3 | 23.8 | 35.0 |
| Method | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | |
| Reference | No Attack Ablation | 29.3 24.9 | 33.4 31.4 | 38.1 31.2 | 40.7 31.6 | 42.1 35.0 | 42.5 34.3 | 45.7 37.5 | 46.9 38.8 | 53.9 48.6 |
| Classification Attack | PGD SI-PGD | 26.4 27.5 | 30.4 31.6 | 34.4 36.1 | 35.4 37.1 | 38.4 | 38.3 | 41.7 | 43.1 | 51.1 |
| MI-DI-PGD | 22.9 | 26.2 | 40.0 | 40.1 | 43.5 | 44.8 | 52.4 | |||
| MI-TI-PGD | 29.3 | 30.0 | 33.2 | 32.1 | 36.0 | 37.5 | 48.0 | |||
| CD-painting | 20.1 16.4 | 23.7 20.8 | 24.9 21.3 | 25.4 22.8 | 30.1 26.6 | 27.4 24.5 | 32.8 28.9 | 34.5 29.5 | 47.1 42.3 | |
| Detection Attack | CD-comics | 16.6 | 21.6 | 21.7 | 22.7 | 26.8 | 24.3 | 29.1 | 42.3 | 43.7 |
| Dfool | 23.3 | 2.5 | 29.2 | 29.8 | 33.3 | 32.9 | 36.5 | 38.0 | 47.5 | |
| Loc DAG | 21.9 20.8 | 0.2 | 25.8 | 26.6 | 29.8 | 29.4 | 33.2 | 33.2 | 45.2 | |
| Ours | RAD | 14.6 | 0.6 0.7 | 22.8 16.3 | 23.4 17.0 | 26.8 20.4 | 25.6 19.1 | 28.9 22.3 | 31.0 23.8 | 40.6 35.0 |
| mAP | mAP50 | mAP75 | |||||||
| Surrogate | M5 | M7 | M8 | M5 | M7 | M8 | M5 | M7 | M8 |
| None | 38.0 | 39.4 | 40.8 | 60.6 | 61.3 | 63.3 | 40.9 | 42.9 | 44.1 |
| Ablation | 31.0 | 31.9 | 33.5 | 51.2 | 51.0 | 53.7 | 32.4 | 34.3 | 35.4 |
| M2 | 17.9 | 18.6 | 20.3 | 31.6 | 31.7 | 34.5 | 18.0 | 18.9 | 20.7 |
| M3 | 11.6 | 11.9 | 12.9 | 19.2 | 19.1 | 20.7 | 12.1 | 12.6 | 13.7 |
| M5 | 1.2 | 11.1 | 11.8 | 2.4 | 17.9 | 18.9 | 1.0 | 11.9 | 12.6 |
| RMSE | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | |
| COCO detection AOCO detection | 0.000 6.469 | 29.3 14.6 | 33.4 0.7 | 38.1 16.3 | 40.7 17 | 42.1 20.4 | 42.5 19.1 | 45.7 22.3 | 46.9 23.8 | 53.9 35.0 |
| COCO segmentation | 0.000 | N | 38.0 | 39.4 | 40.8 | \\\\ | ||||
| AOCO segmentation | 6.606 | 广 | 广 | \\\\ | 1.2 | 一 | 11.1 | 11.8 |
| ID | Model | Backbone | mAP |
| M1 | SSD512 (Liu et al. (2016)) | VGG16 | 29.3 |
| M2 | YOLOv3 (Redmon & Farhadi (2018)) | Darknet | 33.4 |
| M3 | RetinaNet (Lin et al. (2017)) | ResNet-101 | 38.1 |
| M4 | Faster R-CNN (Ren et al. (2015)) | ResNeXt-101-64*4d | 40.7 |
| M5 | Mask R-CNN (He et al. (2017)) | ResNeXt-101-64*4d | 42.1 |
| M6 | Cascade RCNN (Cai & Vasconcelos (2018)) | ResNet-101 | 42.5 |
| M7 | Cascade Mask R-CNN (Cai & Vasconcelos (2018)) | ResNeXt-101-64*4d | 45.7 |
| M8 | Hyrbrid Task Cascade (Chen et al. (2019a)) | ResNeXt-101-64*4d | 46.9 |
| M9 | EfficientDet (Tan et al. (2020)) | EfficientNet+BiFPN | 53.9 |
| Strategy | Parameter | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 |
| Dynamic | -1 | 18.3 | 1.2 | 20.0 | 20.6 | 24.3 | 22.8 | 26.7 | 28.3 | 40.0 |
| -2 | 18.2 | 1.2 | 20.1 | 20.7 | 24.2 | 22.8 | 26.6 | 28.0 | 39.8 | |
| -3 | 18.4 | 1.3 | 20.3 | 20.8 | 24.3 | 22.9 | 26.7 | 28.5 | 40.2 | |
| Static | 10 | 18.2 | 1.1 | 19.9 | 20.5 | 24.2 | 22.8 | 26.2 | 28.0 | 39.8 |
| 20 | 18.1 | 1.2 | 19.9 | 20.5 | 24.3 | 22.6 | 26.4 | 28.2 | 39.9 | |
| 30 | 18.2 | 1.3 | 20.1 | 20.7 | 24.3 | 22.8 | 26.0 | 27.9 | 40.1 |
| Surrogate | M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 |
| YOLOv3 | 14.6 | 0.7 | 16.3 | 17.0 | 20.4 | 19.1 | 22.3 | 23.8 | 35.0 |
| Mask R-CNN | 20.4 | 24.2 | 25.7 | 26.5 | 1.1 | 28.9 | 33.1 | 34.9 | 45.4 |
| RetinaNet | 20.7 | 6.1 | 2.3 | 25.7 | 29.3 | 28.2 | 31.7 | 33.8 | 44.1 |
| M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | |
| COCO detection AOCO detection | 49.2 26.7 | 56.4 1.6 | 58.1 27.6 | 62.0 29.1 | 63.8 34.5 | 60.7 29.9 | 64.1 34.3 | 66.0 37.4 | 74.3 51.7 |
| COCO segmentation | 人 | 一 | 60.6 | 一 | 61.3 | 63.3 | 一 | ||
| AOCO segmentation | 2.4 | 一 | 17.9 | 18.9 | 一 |
| M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | |
| COCO detection AOCO detection | 30.8 14.2 | 35.8 0.6 | 40.6 16.5 | 44.6 17.1 | 46.3 20.8 | 46.3 19.9 | 50.0 23.3 | 51.2 24.6 | 59.9 37.4 |
| COCO segmentation | / | / | 40.9 | 一 | 42.9 | 44.1 | 一 | ||
| AOCO segmentation | 1.0 | / | 11.9 | 12.6 |
| E | Train-Acc. (%) Test-Acc (%) |
| 0.0 99.98 | 95.25 |
| 0.25 | 97.23 92.77 |
| 1.0 86.03 | 81.62 |
| Coarse Class | Fine Classes |
| Dog Bird Insect Monkey Car Feline Truck Fruit Fungus Boat | Chihuahua,Japanese spaniel, Maltese dog,Pekinese, Shih-Tzu, Blenheim spaniel cock,hen,ostrich, brambling,goldfinch, house finch tiger beetle,ladybug,ground beetle, long-horned beetle,leaf beetle, dung beetle guenon, patas, baboon, macaque,langur, colobus jeep,limousine,cab,beach wagon,ambulance,convertible leopard, snow leopard, jaguar, lion,cougar, lynx tow truck, moving van, fire engine, pickup, garbage truck, police van Granny Smith, rapeseed, corn, acorn, hip, buckeye agaric,gyromitra, stinkhorn, earthstar, hen-of-the-woods, coral fungus |
| E | Train-Acc. (%) Test-Acc (%) |
| 0.0 99.98 | 95.25 |
| 0.25 | 97.23 92.77 |
| 1.0 86.03 | 81.62 |
| Coarse Class | Fine Classes |
| Dog Bird Insect Monkey Car Feline Truck Fruit Fungus Boat | Chihuahua,Japanese spaniel, Maltese dog,Pekinese, Shih-Tzu, Blenheim spaniel cock,hen,ostrich, brambling,goldfinch, house finch tiger beetle,ladybug,ground beetle, long-horned beetle,leaf beetle, dung beetle guenon, patas, baboon, macaque,langur, colobus jeep,limousine,cab,beach wagon,ambulance,convertible leopard, snow leopard, jaguar, lion,cougar, lynx tow truck, moving van, fire engine, pickup, garbage truck, police van Granny Smith, rapeseed, corn, acorn, hip, buckeye agaric,gyromitra, stinkhorn, earthstar, hen-of-the-woods, coral fungus |