V7

"--max_distance", "2",
"--max_samples_with_same_target", "5"

Train sequences :

11_700_1200 128_4400_4800 16_0_400 20_400_800 406_100_500 412_900_1300 51_0_400 59_500_900 92_1200_1600 12_1200_1900 147_5400_5800 16_1200_1600 20_800_1200 406_1300_1700 414_0_400 51_400_800 62_0_400 92_400_800 128_10000_11000 147_5800_6200 16_2400_2800 23_500_900 406_4000_4400 415_0_400 51_800_1200 62_400_800 92_800_1200 128_2800_3200 147_6200_6600 20_1200_1600 23_900_1300 406_4400_4800 415_1600_2000 57_400_800 62_800_1200 128_3200_3600 147_9000_9400 20_2000_2400 400_8400_8800 406_900_1300 415_400_800 57_800_1200 74_0_400 128_3600_4000 147_9400_9800 20_2400_2800 400_8800_9200 412_100_500 415_800_1200 59_100_500 74_400_800

Validation sequences:

128_15000_16000 16_3600_4000 400_9200_9600 415_1200_1600 92_1600_2000 128_4000_4400 20_1600_2000 406_1700_2100 57_0_400 92_2000_2400 128_4800_5200 20_4000_4400 406_4800_5200 57_2800_3200 147_10200_10600 20_4400_4800 412_1300_1700 59_900_1300 147_6600_7000 23_2100_2500 412_500_900 74_2800_3200

V8

"--max_distance", "2",
"--max_samples_with_same_target", "5"

Like V7 + oversampled examples at small distances. The way I have oversampled is quite a bit clunky since I still haven't figured out how to sample a desired distribution without the need for hand-tuning the sampling process. Anyway, while creating the prev_valid list I have added the following statement:

if dist < 0.5:
    prev_valid.append(source_frame_index)
     prev_valid.append(source_frame_index)
if dist < 0.3:
     prev_valid.append(source_frame_index)

V9

Moved 57 to the training set completely, because E-99 was disproportionately worse on 57 than other sequences. Decreased max distance to see if 2m was a problem for the expressivity of the network. "--max_distance", "1.3",
"--max_samples_with_same_target", "5"

Train sequences :

11_700_1200 128_4400_4800 16_0_400 20_400_800 406_1300_1700 414_0_400 51_400_800 59_100_500 92_1200_1600 12_1200_1900 147_5400_5800 16_1200_1600 20_800_1200 406_4000_4400 415_0_400 51_800_1200 59_500_900 92_400_800 128_10000_11000 147_5800_6200 16_2400_2800 23_500_900 406_4400_4800 415_1600_2000 57_0_400 62_0_400 92_800_1200 128_2800_3200 147_6200_6600 20_1200_1600 23_900_1300 406_900_1300 415_400_800 57_2800_3200 62_400_800 128_3200_3600 147_9000_9400 20_2000_2400 400_8400_8800 412_100_500 415_800_1200 57_400_800 74_0_400 128_3600_4000 147_9400_9800 20_2400_2800 400_8800_9200 412_900_1300 51_0_400 57_800_1200 74_400_800

Validation sequences:

128_15000_16000 147_10200_10600 20_1600_2000 23_2100_2500 406_1700_2100 412_500_900 62_800_1200 92_2000_2400 128_4000_4400 147_6600_7000 20_4000_4400 400_9200_9600 406_4800_5200 415_1200_1600 74_2800_3200 128_4800_5200 16_3600_4000 20_4400_4800 406_100_500 412_1300_1700 59_900_1300 92_1600_2000

V10

Added a bunch of new data. Note that 128 has overlapping sequences in the training and validation set (128: 10400-10800)

Train sequences:

11_700_1200 128_6400_6800 20_2000_2400 415_0_400 59_500_900 12_1200_1900 128_7200_7600 20_2400_2800 415_1600_2000 62_0_400 127_2000_2400 128_800_1200 20_400_800 415_400_800 62_400_800 127_2800_3200 128_8400_8800 20_800_1200 415_800_1200 64_0_400 127_400_800 128_9200_9600 23_500_900 418_2000_2400 68_400_800 127_800_1200 14_2400_2800 23_900_1300 418_2400_2800 74_0_400 128_10000_11000 14_2800_3000 400_8400_8800 418_400_800 74_400_800 128_10800_11200 147_5400_5800 400_8800_9200 51_0_400 92_1200_1600 128_1200_1600 147_5800_6200 406_1300_1700 51_400_800 92_400_800 128_14400_14800 147_6200_6600 406_4000_4400 51_800_1200 92_800_1200 128_15200_15600 147_9000_9400 406_4400_4800 52_1200_1600 95_1600_2000 128_15600_16000 147_9400_9800 406_900_1300 57_0_400 95_2000_2400 128_2800_3200 16_0_400 412_100_500 57_2800_3200 95_800_1200 128_3200_3600 16_1200_1600 412_900_1300 57_400_800 128_3600_4000 16_2400_2800 414_0_400 57_800_1200 128_4400_4800 20_1200_1600 414_800_920 59_100_500

Validation sequences:

127_0_400 128_4800_5200 20_4000_4400 412_1300_1700 59_900_1300 127_1600_2000 128_6000_6400 20_4400_4800 412_500_900 62_800_1200 128_0_400 128_8000_8400 23_2100_2500 414_0_400 64_400_800 128_10400_10800 147_10200_10600 400_9200_9600 415_1200_1600 74_2800_3200 128_14800_15200 147_6600_7000 406_100_500 418_1200_1600 92_1600_2000 128_15000_16000 16_3600_4000 406_1700_2100 418_2800_3200 92_2000_2400 128_4000_4400 20_1600_2000 406_4800_5200 52_0_400 95_1200_1600

V11

Moved the overlapping 128 sequence to get a new validation set.

V12

Reshuffled train and validation. Set min distance to 0.5

Train sequences:

12_1200_1900 128_15600_16000 14_2400_2800 20_1600_2000 400_8400_8800 412_1300_1700 418_400_800 59_500_900 92_800_1200 127_0_400 128_2800_3200 147_10200_10600 20_2400_2800 400_8800_9200 412_900_1300 51_0_400 62_0_400 95_1200_1600 127_1600_2000 128_3200_3600 147_5400_5800 20_4000_4400 406_100_500 414_0_400 51_400_800 62_400_800 95_1600_2000 127_2000_2400 128_4000_4400 147_6200_6600 20_400_800 406_1700_2100 414_800_920 52_0_400 64_0_400 127_400_800 128_4400_4800 147_6600_7000 21_0_400 406_4000_4400 415_0_400 52_1200_1600 64_400_800 128_0_400 128_6000_6400 147_9400_9800 21_400_800 406_4800_5200 415_1200_1600 57_0_400 74_0_400 128_10000_11000 128_6400_6800 16_0_400 23_2100_2500 406_900_1300 415_2000_2400 57_2800_3200 74_400_800 128_10800_11200 128_8000_8400 16_1200_1600 23_900_1300 408_0_400 415_800_1200 57_400_800 92_1600_2000 128_14800_15200 128_800_1200 16_3600_4000 400_1200_1600 408_800_1200 418_2000_2400 57_800_1200 92_2000_2400 128_15000_16000 128_8400_8800 20_1200_1600 400_400_800 412_100_500 418_2400_2800 59_100_500 92_400_800

Validation sequences:

11_700_1200 128_14400_14800 128_7200_7600 147_9000_9400 20_800_1200 400_2400_2600 408_400_800 418_1200_1600 62_800_1200 95_2000_2400 127_2800_3200 128_15200_15600 128_9200_9600 16_2400_2800 23_2500_2900 400_9200_9600 412_500_900 418_2800_3200 68_400_800 95_800_1200 127_800_1200 128_3600_4000 14_2800_3000 20_2000_2400 23_500_900 406_1300_1700 415_1600_2000 51_800_1200 74_2800_3200 128_1200_1600 128_4800_5200 147_5800_6200 20_4400_4800 400_0_400 406_4400_4800 415_400_800 59_900_1300 92_1200_1600

V13

Added more data.