sobfu
sobfu copied to clipboard
dgrzech
it doesn't converge
at first I removed findMaxUpdateNorm. After a 5 iterations, it goes into inf. energy. Shouldn't the energy be decreasing? And then it crashes at the next data term estimation.
Or if I run the code with findMaxNorm enable, it will crash at 3rd iteration at findMaxNorm
**
**
what data are you running it on? how come regularisation energy isn't 0 at the first iteration of the solver?
I tried disabling the reg-term but it's still not good. Cuda error is at "findMaxUpdatesNorm"
you didn’t answer either of my questions
the cost decreases from iteration 1 to 2, so the solver is working correctly. my guess is that you set the gradient descent step size too high
I want to use your solver in my code to achieve dynamic-fusion like result. So fa, I have implemented a static TSDF library with ICP. Early on, I forgot to initialize the Jacobian matrix while porting your code to my implementation. That's why it did not start at 0. Thank you for pointing that out.
Right now I'm testing your code solely with the depth data captured by myself. However, it never converge even after 2048 iterations at first frame. I tried many configurations including the alpha size and the recommended ones in the paper that you pointed out in your previous response.
I suspect it might be relevant to the noise generated by the depth sensor, but interestingly it does not affect my static fusion.
Interesting... Even though it never converges, the tracking somehow is good enough when there is no more than translations in the data(like you said the solver does work.). The canonical volume can even accumulate in a meaningful way only when it does not contain rotations in the data.
I also tried other step sizes from 0.01 to 0.3. Solver stops working at 0.4
I'm using a 128, 128, 128 grid with 4-10 mm per voxel. Alpha is 0.05 - 0.1 Filter size is 7 lambda is 0.1 weight 0.2
Log:
iter. no. 1 data energy + w_reg * reg energy = 7646.78 + 0.2 * 0 = 7646.78 max. update norm 0.350222 at voxel (87, 48, 61) iter. no. 50 data energy + w_reg * reg energy = 2797.01 + 0.2 * 1.05463e+06 = 213722 max. update norm 0.0398565 at voxel (127, 0, 112) iter. no. 100 data energy + w_reg * reg energy = 2656.16 + 0.2 * 1.30884e+06 = 264424 max. update norm 0.0311408 at voxel (127, 0, 111) iter. no. 150 data energy + w_reg * reg energy = 2599.94 + 0.2 * 1.46987e+06 = 296575 max. update norm 0.0289678 at voxel (127, 0, 110) iter. no. 200 data energy + w_reg * reg energy = 2568.47 + 0.2 * 1.5869e+06 = 319948 max. update norm 0.0297711 at voxel (127, 0, 109) iter. no. 250 data energy + w_reg * reg energy = 2546.59 + 0.2 * 1.67322e+06 = 337190 max. update norm 0.0220492 at voxel (127, 0, 109) iter. no. 300 data energy + w_reg * reg energy = 2530.55 + 0.2 * 1.73912e+06 = 350354 max. update norm 0.0247185 at voxel (127, 0, 108) iter. no. 350 data energy + w_reg * reg energy = 2518.33 + 0.2 * 1.79112e+06 = 360743 max. update norm 0.0193914 at voxel (127, 0, 108) iter. no. 400 data energy + w_reg * reg energy = 2508.63 + 0.2 * 1.83303e+06 = 369114 max. update norm 0.0230706 at voxel (127, 0, 107) iter. no. 450 data energy + w_reg * reg energy = 2500.57 + 0.2 * 1.86783e+06 = 376067 max. update norm 0.0184403 at voxel (127, 0, 107) iter. no. 500 data energy + w_reg * reg energy = 2493.79 + 0.2 * 1.89636e+06 = 381766 max. update norm 0.0155475 at voxel (127, 0, 107) iter. no. 550 data energy + w_reg * reg energy = 2487.84 + 0.2 * 1.92082e+06 = 386651 max. update norm 0.0117325 at voxel (127, 2, 103) iter. no. 600 data energy + w_reg * reg energy = 2483.23 + 0.2 * 1.94245e+06 = 390973 max. update norm 0.0098311 at voxel (127, 2, 103) iter. no. 650 data energy + w_reg * reg energy = 2479.33 + 0.2 * 1.96136e+06 = 394752 max. update norm 0.00923391 at voxel (127, 2, 102) iter. no. 700 data energy + w_reg * reg energy = 2475.37 + 0.2 * 1.97828e+06 = 398130 max. update norm 0.00981192 at voxel (127, 2, 102) iter. no. 750 data energy + w_reg * reg energy = 2471.89 + 0.2 * 1.99309e+06 = 401089 max. update norm 0.00664955 at voxel (127, 2, 102) iter. no. 800 data energy + w_reg * reg energy = 2468.92 + 0.2 * 2.00564e+06 = 403596 max. update norm 0.00483364 at voxel (127, 2, 102) iter. no. 850 data energy + w_reg * reg energy = 2466.35 + 0.2 * 2.01632e+06 = 405731 max. update norm 0.00396784 at voxel (127, 2, 102) iter. no. 900 data energy + w_reg * reg energy = 2464.05 + 0.2 * 2.02548e+06 = 407559 max. update norm 0.00362555 at voxel (127, 2, 102) iter. no. 950 data energy + w_reg * reg energy = 2461.94 + 0.2 * 2.03336e+06 = 409135 max. update norm 0.0032354 at voxel (127, 2, 102) iter. no. 1000 data energy + w_reg * reg energy = 2460.08 + 0.2 * 2.04022e+06 = 410504 max. update norm 0.00412488 at voxel (127, 3, 101) SOLVER REACHED MAX. NO. OF ITERATIONS WITHOUT CONVERGING iter. no. 1 data energy + w_reg * reg energy = 3550.65 + 0.2 * 2.04039e+06 = 411629 max. update norm 0.283898 at voxel (85, 41, 61) iter. no. 50 data energy + w_reg * reg energy = 2231.8 + 0.2 * 2.0562e+06 = 413471 max. update norm 0.0406142 at voxel (127, 0, 104) iter. no. 100 data energy + w_reg * reg energy = 2198.59 + 0.2 * 2.06222e+06 = 414643 max. update norm 0.0362966 at voxel (127, 0, 105) iter. no. 150 data energy + w_reg * reg energy = 2185.54 + 0.2 * 2.06657e+06 = 415499 max. update norm 0.0380663 at voxel (127, 0, 106) iter. no. 200 data energy + w_reg * reg energy = 2177.17 + 0.2 * 2.07041e+06 = 416259 max. update norm 0.0247796 at voxel (127, 0, 107) iter. no. 250 data energy + w_reg * reg energy = 2171.46 + 0.2 * 2.0736e+06 = 416891 max. update norm 0.0132635 at voxel (127, 0, 109) iter. no. 300 data energy + w_reg * reg energy = 2166.9 + 0.2 * 2.07631e+06 = 417429 max. update norm 0.00937067 at voxel (127, 0, 110) iter. no. 350 data energy + w_reg * reg energy = 2163.27 + 0.2 * 2.07877e+06 = 417917 max. update norm 0.00749873 at voxel (127, 4, 108) iter. no. 400 data energy + w_reg * reg energy = 2160.23 + 0.2 * 2.08057e+06 = 418274 max. update norm 0.00592503 at voxel (127, 0, 111) iter. no. 450 data energy + w_reg * reg energy = 2157.65 + 0.2 * 2.08197e+06 = 418551 max. update norm 0.00498483 at voxel (127, 5, 107) iter. no. 500 data energy + w_reg * reg energy = 2155.44 + 0.2 * 2.08312e+06 = 418779 max. update norm 0.00444591 at voxel (127, 5, 107) iter. no. 550 data energy + w_reg * reg energy = 2153.42 + 0.2 * 2.08412e+06 = 418978 max. update norm 0.00403678 at voxel (127, 5, 107) iter. no. 600 data energy + w_reg * reg energy = 2151.44 + 0.2 * 2.08464e+06 = 419080 max. update norm 0.00387362 at voxel (127, 4, 109) iter. no. 650 data energy + w_reg * reg energy = 2149.45 + 0.2 * 2.08471e+06 = 419091 max. update norm 0.00376256 at voxel (127, 4, 109) iter. no. 700 data energy + w_reg * reg energy = 2147.63 + 0.2 * 2.08478e+06 = 419103 max. update norm 0.0033257 at voxel (127, 4, 109) iter. no. 750 data energy + w_reg * reg energy = 2146.04 + 0.2 * 2.08474e+06 = 419093 max. update norm 0.00286596 at voxel (127, 4, 109) iter. no. 800 data energy + w_reg * reg energy = 2144.6 + 0.2 * 2.08459e+06 = 419062 max. update norm 0.00256356 at voxel (127, 4, 109) iter. no. 850 data energy + w_reg * reg energy = 2143.31 + 0.2 * 2.08465e+06 = 419073 max. update norm 0.00232556 at voxel (127, 4, 109) iter. no. 900 data energy + w_reg * reg energy = 2142.09 + 0.2 * 2.08484e+06 = 419110 max. update norm 0.00211953 at voxel (127, 5, 108) iter. no. 950 data energy + w_reg * reg energy = 2140.93 + 0.2 * 2.08484e+06 = 419108 max. update norm 0.00199594 at voxel (127, 5, 108) iter. no. 1000 data energy + w_reg * reg energy = 2139.8 + 0.2 * 2.08468e+06 = 419076 max. update norm 0.00183652 at voxel (127, 5, 108) SOLVER REACHED MAX. NO. OF ITERATIONS WITHOUT CONVERGING iter. no. 1 data energy + w_reg * reg energy = 6906.06 + 0.2 * 2.08467e+06 = 423839 max. update norm 0.510061 at voxel (88, 46, 63) iter. no. 50 data energy + w_reg * reg energy = 2941.66 + 0.2 * 2.08882e+06 = 420706 max. update norm 0.0442168 at voxel (41, 78, 120) iter. no. 100 data energy + w_reg * reg energy = 2714.22 + 0.2 * 2.08955e+06 = 420625 max. update norm 0.0231331 at voxel (39, 78, 127) iter. no. 150 data energy + w_reg * reg energy = 2639.57 + 0.2 * 2.0898e+06 = 420599 max. update norm 0.0112255 at voxel (127, 2, 110) iter. no. 200 data energy + w_reg * reg energy = 2600.44 + 0.2 * 2.08997e+06 = 420594 max. update norm 0.0100529 at voxel (127, 0, 110) iter. no. 250 data energy + w_reg * reg energy = 2576.55 + 0.2 * 2.0899e+06 = 420556 max. update norm 0.01044 at voxel (127, 0, 110) iter. no. 300 data energy + w_reg * reg energy = 2559.94 + 0.2 * 2.09032e+06 = 420623 max. update norm 0.00768468 at voxel (127, 0, 109)
alpha >= 0.1 will definitely be too large. the reconstructions on youtube were run with a step size of the order of 1e-4. for more on convergence see the discussion in #2
re translation/rotation the problem with rotations would largely be due to the fact that, in the presence of rotation, the canonical tsdf is too different from the live projective sdf. the entire deformation field is estimated starting at frame 0, meaning that for parts of the volume that only come into view later on, the field will be very inaccurate, which it's impossible to fix, and the problem only gets worse at subsequent frames. you can see the same issue in the snoopy video on youtube