File size: 14,054 Bytes
a3778d1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 | ====================================================================== Model: molmo-7B-O-0924 ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 338 363 93.1% Depth counter 49 65 75.4% Depth ambiguous 120 172 69.8% ------------------------------------------------------ TOTAL consistent 338 363 93.1% TOTAL counter 49 65 75.4% Accuracy Gap (Consistent - Counter): 17.7%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 16 / 65 ====================================================================== Model: molmo-7B-O-0924-data_scale_exp_80k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 291 363 80.2% Depth counter 37 65 56.9% Depth ambiguous 98 172 57.0% ------------------------------------------------------ TOTAL consistent 291 363 80.2% TOTAL counter 37 65 56.9% Accuracy Gap (Consistent - Counter): 23.2%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 28 / 65 ====================================================================== Model: molmo-7B-O-0924-data_scale_exp_400k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 325 363 89.5% Depth counter 37 65 56.9% Depth ambiguous 118 172 68.6% ------------------------------------------------------ TOTAL consistent 325 363 89.5% TOTAL counter 37 65 56.9% Accuracy Gap (Consistent - Counter): 32.6%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 28 / 65 ====================================================================== Model: molmo-7B-O-0924-data_scale_exp_800k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 322 363 88.7% Depth counter 46 65 70.8% Depth ambiguous 124 172 72.1% ------------------------------------------------------ TOTAL consistent 322 363 88.7% TOTAL counter 46 65 70.8% Accuracy Gap (Consistent - Counter): 17.9%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 19 / 65 ====================================================================== Model: molmo-7B-O-0924-data_scale_exp_2m ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 329 363 90.6% Depth counter 47 65 72.3% Depth ambiguous 148 172 86.0% ------------------------------------------------------ TOTAL consistent 329 363 90.6% TOTAL counter 47 65 72.3% Accuracy Gap (Consistent - Counter): 18.3%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 18 / 65 ====================================================================== Model: NVILA-Lite-2B ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 270 363 74.4% Depth counter 26 65 40.0% Depth ambiguous 119 172 69.2% ------------------------------------------------------ TOTAL consistent 270 363 74.4% TOTAL counter 26 65 40.0% Accuracy Gap (Consistent - Counter): 34.4%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 39 / 65 ====================================================================== Model: NVILA-Lite-2B-data-scale-exp-80k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 260 363 71.6% Depth counter 33 65 50.8% Depth ambiguous 104 172 60.5% ------------------------------------------------------ TOTAL consistent 260 363 71.6% TOTAL counter 33 65 50.8% Accuracy Gap (Consistent - Counter): 20.9%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 32 / 65 ====================================================================== Model: NVILA-Lite-2B-data-scale-exp-400k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 295 363 81.3% Depth counter 38 65 58.5% Depth ambiguous 113 172 65.7% ------------------------------------------------------ TOTAL consistent 295 363 81.3% TOTAL counter 38 65 58.5% Accuracy Gap (Consistent - Counter): 22.8%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 27 / 65 ====================================================================== Model: NVILA-Lite-2B-data-scale-exp-800k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 307 363 84.6% Depth counter 44 65 67.7% Depth ambiguous 118 172 68.6% ------------------------------------------------------ TOTAL consistent 307 363 84.6% TOTAL counter 44 65 67.7% Accuracy Gap (Consistent - Counter): 16.9%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 21 / 65 ====================================================================== Model: NVILA-Lite-2B-data-scale-exp-2m ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 353 363 97.2% Depth counter 61 65 93.8% Depth ambiguous 149 172 86.6% ------------------------------------------------------ TOTAL consistent 353 363 97.2% TOTAL counter 61 65 93.8% Accuracy Gap (Consistent - Counter): 3.4%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 4 / 65 ====================================================================== Model: RoboRefer-2B-SFT ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 359 363 98.9% Depth counter 62 65 95.4% Depth ambiguous 153 172 89.0% ------------------------------------------------------ TOTAL consistent 359 363 98.9% TOTAL counter 62 65 95.4% Accuracy Gap (Consistent - Counter): 3.5%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 3 / 65 ====================================================================== Model: Qwen2.5-VL-3B-Instruct ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 274 363 75.5% Depth counter 36 65 55.4% Depth ambiguous 112 172 65.1% ------------------------------------------------------ TOTAL consistent 274 363 75.5% TOTAL counter 36 65 55.4% Accuracy Gap (Consistent - Counter): 20.1%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 29 / 65 ====================================================================== Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_80k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 253 363 69.7% Depth counter 39 65 60.0% Depth ambiguous 96 172 55.8% ------------------------------------------------------ TOTAL consistent 253 363 69.7% TOTAL counter 39 65 60.0% Accuracy Gap (Consistent - Counter): 9.7%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 26 / 65 ====================================================================== Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_400k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 239 363 65.8% Depth counter 38 65 58.5% Depth ambiguous 95 172 55.2% ------------------------------------------------------ TOTAL consistent 239 363 65.8% TOTAL counter 38 65 58.5% Accuracy Gap (Consistent - Counter): 7.4%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 27 / 65 ====================================================================== Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_800k ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 222 363 61.2% Depth counter 38 65 58.5% Depth ambiguous 92 172 53.5% ------------------------------------------------------ TOTAL consistent 222 363 61.2% TOTAL counter 38 65 58.5% Accuracy Gap (Consistent - Counter): 2.7%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 27 / 65 ====================================================================== Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_2m ====================================================================== Category Type Correct Total Accuracy ------------------------------------------------------ Depth consistent 225 363 62.0% Depth counter 35 65 53.8% Depth ambiguous 91 172 52.9% ------------------------------------------------------ TOTAL consistent 225 363 62.0% TOTAL counter 35 65 53.8% Accuracy Gap (Consistent - Counter): 8.1%p -> Larger gap indicates stronger reliance on the 2D heuristic 🔍 Counter examples wrong: 30 / 65 ============================================================================== MODEL COMPARISON (CV-Bench-3D depth) ============================================================================== Model Consistent Counter Gap ------------------------------------------------------------------------------ molmo-7B-O-0924 93.1% 75.4% +17.7%p molmo-7B-O-0924-data_scale_exp_80k 80.2% 56.9% +23.2%p molmo-7B-O-0924-data_scale_exp_400k 89.5% 56.9% +32.6%p molmo-7B-O-0924-data_scale_exp_800k 88.7% 70.8% +17.9%p molmo-7B-O-0924-data_scale_exp_2m 90.6% 72.3% +18.3%p ------------------------------------------------------------------------------ NVILA-Lite-2B 74.4% 40.0% +34.4%p NVILA-Lite-2B-data-scale-exp-80k 71.6% 50.8% +20.9%p NVILA-Lite-2B-data-scale-exp-400k 81.3% 58.5% +22.8%p NVILA-Lite-2B-data-scale-exp-800k 84.6% 67.7% +16.9%p NVILA-Lite-2B-data-scale-exp-2m 97.2% 93.8% +3.4%p RoboRefer-2B-SFT 98.9% 95.4% +3.5%p ------------------------------------------------------------------------------ Qwen2.5-VL-3B-Instruct 75.5% 55.4% +20.1%p Qwen2.5-VL-3B-Instruct-data_scale_exp_80k 69.7% 60.0% +9.7%p Qwen2.5-VL-3B-Instruct-data_scale_exp_400k 65.8% 58.5% +7.4%p Qwen2.5-VL-3B-Instruct-data_scale_exp_800k 61.2% 58.5% +2.7%p Qwen2.5-VL-3B-Instruct-data_scale_exp_2m 62.0% 53.8% +8.1%p |