Our method uses the ground truth segmentation image in training, but debugging results seem to indicate that this data is not actually coming through properly in the images. How can we fix this?
Check your active sensors in config.py, afaik the order matters but I would check that assumption because list. Since we aren’t getting a dictionary it’s difficult to be certain but when I was previously checking input sizes they seemed to be fine.
I would also check the number of images. Obviously for single camera it should be 2 and you would expect segmentation mask to be index 1, might pay to check if 6 exist as per multi camera.
Hey, those are good points. We’ve tested to make sure that the images we see are in the right format, and in the right number. We tested the BEV segmentation image locally and printed some debug statements on the official platform, and get wildly different results. The following is the output of the green channel in the online platform:
bev_camera row [202 202 202 202 202 200 203 202 202 204 205 202 193 199 204 207 208 206
204 203 206 206 206 206 208 208 208 206 207 207 208 208 208 209 209 209
209 210 211 214 211 206 205 203 205 205 190 178 184 192 196 196 197 202
201 200 199 199 202 206 205 203 200 206 206 206 206 205 205 206 209 209
206 205 202 202 205 203 202 203 202 202 205 205 204 205 207 209 211 209
208 205 208 208 208 208 209 210 211 212 212 209 209 208 206 204 207 209
210 203 201 196 198 201 211 219 229 235 227 218 208 202 198 202 203 203
208 210 210 211 214 212 208 210 211 211 211 211 211 211 208 204 213 222
229 218 208 205 211 212 212 213 213 213 212 210 210 211 211 213 214 213
211 210 213 211 213 210 208 210 213 215 215 215 214 212 210 211 214 213
212 213 214 214 214 211 209 208 207 209 217 218 213 215 220 217 213 211
206 205 206 203 202 205 199 186 189 197 185 188 208 218 237 237 217 200
201 202 200 200 198 197 197 196 193 190 188 186 188 186 186 186 185 185
187 186 186 185 185 187 189 190 191 193 194 196 194 196 199 199 196 196
196 197 197 196 197 196 195 194 192 194 193 190 189 189 189 189 188 188
186 187 187 188 190 190 190 191 192 190 190 192 195 196 197 217 237 237
219 221 236 222 198 185 197 205 206 196 181 196 204 203 197 186 189 200
203 206 205 200 202 203 205 206 205 206 206 204 205 207 205 203 205 210
209 205 206 205 204 205 207 208 210 208 205 204 204 202 198 197 198 199
198 194 197 206 214 224 228 232 226 217 208 201 198 197 204 208 214 209
208 209 208 208 209 209 209 208 206 210 213 210 208 210 211 212 215 218
219 220 220 219 217 212 206 206 208 209 209 211 213 212 211 211 211 210
208 208 211 213 212 211 210 211 210 210 209 210 211 213 211 210 211 210
206 206 209 209 206 202 196 191 195 198 200 204 219 216 232 235 231 215
187 182 187 195 206 209 208 197 191 192 195 194 191 189 190 190 186 184
179 182 180 173 176 181 182 183 184 188 191 197 200 199 194 189 189 190
190 189 183 178 179 188 199 200 201 200 197 193 193 194 199 202 201 202
203 203 204 203 200 199 196 193 193 196 195 190 186 183 178 181 175 185
203 215 221 213 213 206 199 194]
And the same locally:
bev_camera row [255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 80 80 80 80 80 80 80 80 80 80 80 80
80 80 80 80 80 80 80 80 80 80 80 80 80 80 80 80 80 80
80 80 80 80 80 80 80 80 80 80 80 80 80 80 80 80 80 80
80 80 80 80 80 80 80 80 80 80 80 80 80 80 80 80 80 80
80 80 80 80 80 80 80 80 80 80 80 80 80 80 80 80 80 80
80 80 80 80 80 80 80 80 80 80 80 80 80 80 80 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255 255
255 255 255 255 255 255 255 255]
We also printed the other (front camera), and confirmed that it also does not match the pattern we expect.
Unfortunately having the agent running on local doesn’t mean it will run on the challenge server (re: my 20+ failed debugging submissions over the last month). Keep in mind that there is a significant amount of files overwritten for the test phase so changes to some files don’t result in changes during testing. I’m not sure what issue you are encountering, sorry I can’t be of any more help.
Hey, I was hoping to get an update from the organizers on this. Our team spent significant time and effort developing our approaches, and it would be great to understand either what went wrong, or potentially rerun our existing submissions on a repaired pipeline.